In this paper
Key findings
- The attacks disclosed that week came through access held by vendors: roles left in place for a former data vendor (Wiz, 9 November 2026), a key issued to a company that runs AI models (Cursor, 13 November 2026) and a key never retired after a project (Barracuda, 11 November 2026).
- The FBI's FLASH notice of 13 November 2026 says some intrusions built on the ARTEX agent framework reached privileged access in under three hours, and it recommends zero trust access.
- All three clouds now give an agent its own identity. Of their documentation read on 17 November 2026, only the Azure pages show a single switch that turns off model API keys, and Microsoft warns that old keys can keep working for up to several hours.
- Authorising each tool call has been generally available on AWS since 3 March 2026, while Microsoft's tool policies for Foundry agents are preview and Google's pages disagree on the status of its per-tool rules.
- Full logging of model requests and responses is off by default on AWS, and the Google Cloud audit logs that record model calls are off by default.
- Anthropic and OpenAI both say in writing that prompt injection, where instructions are hidden in what a model reads, is not fully solved, so this design keeps its controls outside the model.
Summary
Zero trust means that no request is trusted because of where it comes from. This paper applies that idea to AI models, to the data they read and to the agents that act with a person's access. It starts from the verified security events of 7 to 17 November 2026: attacks that came through keys and roles held by vendors, AI vendors among them, and through agents and their tools. It then sets out a reference design with seven control points, shows how to build it on Microsoft Azure, AWS and Google Cloud and with Anthropic and OpenAI, and gives a 40-point checklist with the evidence for each check and an order for the first 30 days.
Designed PDF
Download the PDF
The full paper is free to read on this page. The designed PDF arrives in your inbox, attached, within a minute.
Section 01Summary: what zero trust means for AI, and what we found
Zero trust means that no request is trusted because of where it comes from. Every caller, whether a person, a program or an AI agent, has to prove who it is on each request. Each request gets the least access the task needs and leaves a record. NIST set this out in Special Publication 800-207 in August 2020. Applying it to AI means treating the model and the agent as callers like any other, and treating everything they read as untrusted.
The week of 7 to 17 November 2026 is the reason to do this now. The attacks disclosed that week came through keys and roles that companies had handed to vendors, AI vendors among them, and through AI agents and the tools they call. In each case the attacker arrived with access that somebody had granted on purpose.
In the vendors' documentation we found that the parts exist on all three large clouds and at both model makers, and that many of the parts closest to agents are still preview. Each cloud now gives an agent its own identity, a private path to the model and an inspection service for prompts and outputs. Full logging of model calls is often off until someone turns it on. For authorising each tool call, AWS's policy engine is generally available, Microsoft's tool policies for Foundry agents are preview, and Google's pages disagree on the status of its per-tool rules.
A zero trust design for AI asks seven questions of every request: who is calling, what it may reach, which path it takes, which tool it uses, what content goes in and out, which data it touches, and what record is left. The checklist in this paper turns them into 40 checks, each with the evidence that proves it.
Section 02Why now: the week of 7 to 17 November 2026
Three kinds of event filled the week: attacks through access held by vendors, attacks aimed at the users of a workplace email service, and AI security disclosures. We list only what a company, a government body or a named research firm confirmed. Claims by attackers that no victim confirmed are left out.
Start with email. Microsoft's security updates of 10 November 2026 fixed CVE-2026-124194, an Outlook flaw that Microsoft lists as already exploited and that exposes a user's password hash. On 13 November 2026 a high-volume campaign sent fake Microsoft 365 service notices, which Microsoft's security blog of 16 November 2026 rates as likely credential phishing. Both aimed at the same prize as the vendor cases below: a credential that the rest of the system then trusts.
| Event | Source and date | What it was | What was trusted by default |
|---|---|---|---|
| Fake Microsoft 365 service notices by email, 13 November | Microsoft Security Blog, 16 November 2026 | Likely credential phishing, according to Microsoft. | A message that looked like the vendor's own notice. |
| Outlook flaw CVE-2026-124194, patched 10 November | Microsoft Security Response Center, 10 November 2026 | Exploited vulnerability that exposes a user's password hash. | An outbound connection made by a trusted mail client. |
| Roles created for the data vendor Qubole, used 14 to 26 October | Wiz, 9 November 2026. SecurityWeek, 10 November 2026 | Attack with the vendor's own AWS credentials at more than a dozen organisations. | A role left in place after the customer stopped using the vendor. |
| A key Cursor had issued to Fireworks AI, attacker activity 29 October to 1 November | Cursor, 13 November 2026 | Attack through a company that runs AI models for others. Malicious copies of a coding agent extension sat in Cursor's marketplace cache for about 11 hours. | An access key held by a supplier. |
| A cloud key Barracuda had given to an AI vendor, used 1 to 2 November | Barracuda, 11 November 2026 | Attack with a valid key that crossed about 10 cloud accounts in about 44 hours. | A key from a finished project that nobody retired. |
| ARTEX, an open-source AI agent framework used by attackers | FBI FLASH-20261113-001, 13 November 2026 | Government warning. Some intrusions reached privileged access in under three hours. | Leftover credentials, partner tenants and app consent grants. |
| MCP server flaws CVE-2026-126681 and CVE-2026-109492 | AWS security bulletin, 16 November 2026. Microsoft security release, 10 November 2026 | Disclosed vulnerabilities. AWS's bulletin does not report exploitation. | Text read by an agent's tool, and other users of a shared gateway. |
The three vendor attacks in the table have one thing in common. Each used access that a company had granted to a supplier on purpose, so the attacker arrived with valid credentials. Wiz reported on 9 November 2026 that in most environments it examined, the attacker's sign-ins were the only use of the Qubole role that month.
Speed is the second thing, and the FBI's figure of under three hours is the one to plan against. Its advice includes moving enterprise applications behind zero trust access and applying least privilege to app consent grants. The last row concerns MCP, the Model Context Protocol, which is the common way an agent connects to tools. In the case AWS disclosed on 16 November 2026, injected text could make an MCP server send its own database token to an attacker.
Section 03Zero trust in plain words, and what changes when the caller is a model or an agent
Zero trust is a way of deciding access. A request gets no credit for coming from inside the office network or from a known vendor. It has to prove who is asking each time (security people call this continuous verification), and it gets only the access the task needs (least privilege). The design assumes that something inside is already compromised.
NIST SP 800-207 sets out seven tenets. In plain words they are these.
- Every data source and computing service is a resource to protect.
- All communication is secured, wherever on the network it happens.
- Access is granted one session at a time, with the least privilege needed.
- Access is decided by a policy that looks at the caller's identity, the application, the device and their behaviour.
- The organisation watches the security state of everything it owns and trusts none of it by default.
- Authentication and authorisation happen before access, every time.
- The organisation collects what it can about assets and traffic and uses it to improve security.
CISA's Zero Trust Maturity Model, version 2.0 (cover dated April 2023), sorts the work into five pillars: identity, devices, networks, applications and workloads, and data. Three capabilities run across all five: visibility and analytics, automation and orchestration, and governance. The fourth NIST tenet already allows the caller to be a service identity as well as a person, and a NIST concept paper of 5 February 2026 on agent identity says it will apply SP 800-207 to agents. Three things change when the caller is a model or an agent.
- It reads untrusted text. A model cannot reliably tell its owner's instruction from one hidden in a web page, an email or a tool result. The hidden kind is called indirect prompt injection, and prompt injection tops OWASP's list of risks for language model applications.
- It acts with someone's access. Whatever the agent's identity can reach, a hijacked agent can reach. OWASP calls giving an agent more than it needs excessive agency.
- It chains tools. One step's output becomes the next step's input, so one bad input can cross several systems before a person sees anything. OWASP's list for agentic applications of 9 December 2025 names tool misuse and cascading failures.
The consequence for design is that the agent cannot be where the rules are kept. A rule written into the prompt can be overridden by other text the model reads. The controls that hold sit outside the model: the identity it runs as, the network it sits in, the gateway its tool calls pass through, and the log it cannot edit.
Section 04The reference design: seven control points every request passes
The design is a path with seven checkpoints. A request starts with a person or a scheduled job, passes through an agent, reaches a model, and often goes on to tools and data. At each checkpoint a control outside the model asks one question and can refuse.
- 1. Who is calling. Every person, application and agent has its own identity, and nothing calls a model with a shared or long-lived key.
- 2. What it may reach. A policy decides, for this caller and this request, which models, tools and data are allowed. The default answer is no.
- 3. The path. Traffic to the model and to tools uses a private route, and every route that leaves it is known.
- 4. The tool. Each tool call passes a gateway that checks an allow list, and actions that cannot be undone wait for a person.
- 5. The content in and out. Prompts, documents, tool results and outputs are inspected for injected instructions and sensitive data.
- 6. The data. The agent sees only what the person it works for may see, and the company knows where prompts go and how long they are kept.
- 7. The record. Every model call and tool call is logged with the identities involved, and detection runs on those logs.
The seven control points on the path of one AI request
The policy engine sits above the flow because it is set once for the organisation and applied at every gate, which is how a rule such as approved models only holds in every team's account. Vendors are drawn as ordinary callers because the week's attacks came through access that vendors already held.
Status labels in the next four sections are the vendors' own, read on the morning of 17 November 2026. Microsoft Ignite opened that day, and this paper reflects the documentation as it stood before any announcement there.
Section 05Building it on Microsoft Azure
On Microsoft's cloud the model service is Microsoft Foundry and the identity service is Microsoft Entra. Start with identity, because several later controls work only when the caller holds an Entra token. Give each agent an Entra Agent ID, move applications to managed identities (identities that Azure issues and renews, with no secret to store), and turn off key access on every Foundry resource.
| Control point | Product | Status | Limit to know |
|---|---|---|---|
| 1. Who is calling | Entra Agent ID. Managed identities for applications. | Generally available since April 2026. Some features, such as agents that run on a user's device, are preview. | Entra security features for agents need an Agent 365 licence. Legacy Foundry agents share one identity per project. |
| 2. What it may reach | Conditional Access for agents. The disableLocalAuth setting with Azure Policy. The approved models policy. | The key switch and model policy are generally available. Microsoft's pages disagree on whether Conditional Access for agents is preview. | Policy is checked when a token is issued, and API key calls bypass it. Key cutoff can take several hours. |
| 3. The path | Private Link, the network security perimeter, and private networking for Agent Service. | Private Link is generally available. The perimeter is generally available for Foundry resources and preview for the Azure OpenAI kind. | Tools that fetch web or SharePoint content (Bing grounding, web search, SharePoint grounding) still use public endpoints. |
| 4. The tool | Azure Policy for approved and denied tools, the AI gateway for MCP tools, and API Management. | The tool policies and AI gateway governance are preview. API Management's MCP features carry no status label. | API Management applies policy to a whole MCP server. Toolbox approval flags are enforced only by the software that runs the agent. |
| 5. The content in and out | Prompt Shields and Foundry guardrails. | Prompt Shields carries no preview label. Agent guardrails and tool response checks are preview. | Document detection is off by default. Prompt Shields is trained mainly on English. Guardrails apply only to Models sold by Azure. |
| 6. The data | Microsoft Purview and customer-managed keys. | Generally available. | Purview policies apply only to calls that carry user context. No abuse monitoring retention period in days is stated. |
| 7. The record | Diagnostic settings, agent tracing, and Defender for Cloud AI threat protection. | Generally available. Tracing is preview for workflow and external agents. | Threat protection scans text only. Microsoft's page does not say whether diagnostic logs hold prompt text. |
Much of the tool layer on Azure is preview. The controls documented for Models sold by Azure, such as Foundry guardrails and Defender's AI threat protection, are not documented for Claude models, which Microsoft lists as a marketplace offer that Anthropic operates.
Section 06Building it on AWS
On AWS the model service is Amazon Bedrock and the agent platform is Amazon Bedrock AgentCore. The pages we read show no single switch for Bedrock API keys, so start by denying them in both the bedrock and the bedrock-mantle namespaces, because denying only the first leaves the second endpoint open. Then turn on model invocation logging, which is off by default.
| Control point | Product | Status | Limit to know |
|---|---|---|---|
| 1. Who is calling | AgentCore Identity workload identities. IAM roles with temporary credentials from STS. | Generally available, with no preview label on the pages. Token exchange on behalf of a user was added in April 2026. | GetWorkloadAccessTokenForUserId accepts a user ID the caller supplies. AWS recommends denying it in production. |
| 2. What it may reach | IAM condition keys, organisation-wide service control policies, and AgentCore's check on signed tokens (the JWT authorizer). | Generally available. | Short-term keys cannot be revoked one by one. The pages we read document no product named conditional access for agents. |
| 3. The path | PrivateLink for Bedrock and AgentCore. VPC mode for AgentCore Runtime. | Generally available. | The default endpoint policy allows full access. Endpoint policies restrict IAM principals and cannot restrict OAuth users. |
| 4. The tool | AgentCore Gateway, Policy in AgentCore (written in the Cedar language), and interceptors. | Policy has been generally available since 3 March 2026, in 13 Regions. | The gateway's shared role is the ceiling for every caller, so AWS advises one gateway per trust boundary. |
| 5. The content in and out | Bedrock Guardrails and the ApplyGuardrail API. | Generally available. | The prompt attack filter does not evaluate tool results or tool definitions. Guardrails do not run on bedrock-mantle. |
| 6. The data | Data retention modes, customer encryption keys in KMS, and retrieval that respects access lists. | Generally available. | Retention mode is set by API only. Claude Fable 5 models require up to 30 days retention for most customers, with possible AWS staff review. |
| 7. The record | Model invocation logging, CloudTrail, AgentCore Observability and GuardDuty AI Protection. | Logging is generally available. GuardDuty AI Protection launched on 14 July 2026. Its launch post carries no status label, and the Security Hub blog calls it generally available. | Invocation logging is off by default and covers bedrock-runtime only. |
The bedrock-mantle endpoint is the main thing to watch, since Guardrails and invocation logging do not run on it. AWS's pages also disagree on long-term API keys: the API keys page calls them suitable for exploration only, while sample code on two model cards tells readers to create one.
Section 07Building it on Google Cloud
Google's AI service is Gemini Enterprise Agent Platform (called Vertex AI until 22 April 2026), and its managed home for agents is Agent Runtime (formerly Vertex AI Agent Engine). Two defaults need attention first. Agent Platform's public endpoints stay reachable from the internet until a VPC Service Controls perimeter is in place, and an agent on Agent Runtime runs as an ordinary service account unless the team opts in to Agent Identity.
| Control point | Product | Status | Limit to know |
|---|---|---|---|
| 1. Who is calling | Agent Identity, with a certificate that lasts 24 hours and renews itself. | Generally available according to the IAM release notes of 22 April 2026. Tokens for other clouds and MCP servers are preview. | Opt-in on Agent Runtime. Redeploying an agent creates a new principal, so earlier grants do not carry over. |
| 2. What it may reach | IAM Conditions, deny policies, principal access boundary (an outer limit on what an identity can reach), and organisation policy on keys and models. | Generally available. The status of principal access boundary for agent identities is not stated. | The key constraint allows the Gemini API by default. The model policy covers Model Garden models only. |
| 3. The path | VPC Service Controls with Private Service Connect. | Generally available. | Private Service Connect alone leaves the public route open. Inside a perimeter, request-response logging is not available. |
| 4. The tool | Agent Gateway with Agent Registry, Unified Access Policies, and semantic governance, in which a model judges each tool call. | Agent Gateway has been generally available since 18 June 2026. Semantic governance is preview. Google's pages disagree on Unified Access Policies. | A wildcard rule skips outbound governance. Google says semantic governance verdicts may not be accurate. |
| 5. The content in and out | Model Armor. | Generally available on Agent Gateway since 24 June 2026. | The built-in Agent Platform integration continues without screening when Model Armor is unreachable. Model Armor checks one prompt or response at a time. |
| 6. The data | Customer-managed encryption keys, Sensitive Data Protection, and permission-aware Agent Search. | Keys are listed model by model. Permission-aware search is preview. | Google says security controls are not supported for preview models. |
| 7. The record | Cloud Audit Logs, inference metadata logging, and Security Command Center AI Protection. | Metadata logging has been generally available since 30 October 2026. AI Protection is generally available in the Premium tier, with agent workload support in preview. | Data Access logs, which record model calls, are off by default. |
Unified Access Policies are the rules that allow or deny each tool call at the gateway. Google's policy overview and its IAM release notes of 2 November 2026 call them preview, and the Agent Platform release notes of 31 August 2026 call them generally available.
Section 08Building it with the model makers directly: Anthropic and OpenAI
A company can also buy the model from its maker, through the API or a workplace app such as Claude Enterprise or ChatGPT Enterprise. The table sets Anthropic and OpenAI side by side on the seven control points, with the status where one has a control the other lacks.
| Control point | Anthropic | OpenAI |
|---|---|---|
| 1. Who is calling | Single sign-on. Service-account keys, and Workload Identity Federation, which swaps a workload's own signed identity for a short-lived token. Static keys cannot be limited to specific actions. | Single sign-on. Project service accounts, keys that can be scoped down, and workload identity federation that also accepts X.509 certificates. |
| 2. What it may reach | Workspaces with their own members, keys and limits. An organisation policy caps key expiry. | Projects with their own roles. Admins can cap key lifetime and allow or deny models and hosted tools per project. |
| 3. The path | IP allowlist in beta. Inference geography is global or US. | IP allowlist for all API customers. Mutual TLS (a client certificate check), which the developer documentation calls generally available and a help page calls beta. Azure Private Link through sales. |
| 4. The tool | On Enterprise, each connector or tool can be allowed, sent for approval or blocked. Claude Code takes managed settings, and its sandbox is off by default. | In new Enterprise workspaces apps stay disabled until an admin enables them. Codex takes managed configuration, and its sandbox has network access off by default. |
| 5. The content in and out | Inference hooks on Enterprise, generally available since 20 October 2026, send each prompt to the customer's own security server for a decision. | Agents SDK guardrails on input, output and tools. Lockdown Mode limits outbound requests after an injection. |
| 6. The data | On the API, 30 days retention as standard (the Privacy Center figure). Zero retention by contract, excluding several agent and storage features. The newest Fable and Mythos models require 30 days. | On the API, abuse monitoring logs kept up to 30 days. Zero retention by approval, excluding stateful features such as Files and Batch. In-region processing in six regions for eligible organisations. |
| 7. The record | Compliance API with an activity feed kept 6 years. Enterprise audit export for 180 days. | Audit logs API for admin actions, with no per-request inference log. Compliance Platform for Enterprise and Edu. |
Reaching the same model through a cloud changes who is responsible. Identity and network controls become the cloud's own, and the model maker's admin tools mostly drop away. Anthropic's pages list its Admin and Compliance APIs as not supported on Amazon Bedrock or on the version Google runs.
Each cloud now carries two different Claude products. On AWS and Google Cloud, one is run by the cloud itself (Claude on Amazon Bedrock, Claude run by Google on Agent Platform) and the other is run by Anthropic and sold through the cloud (Claude Platform on AWS, the service at claude.googleapis.com). On the kind Anthropic runs, data can be processed outside the cloud, and most of that cloud's controls stop applying. Microsoft Foundry also has two options, hosted on Azure or on Anthropic infrastructure, and Anthropic operates both as an independent processor.
The companies describe the Azure-hosted Claude option differently: Microsoft's page says prompts and outputs stay in Azure, Anthropic's Foundry page says operational data is stored outside Azure, and Anthropic's Claude Code page says flagged content goes to Anthropic. OpenAI's models also differ by cloud. On Azure, Microsoft hosts them with no calls to OpenAI's services, and on Amazon Bedrock, OpenAI's hosted API is not in the request path. On Google Cloud they were in private preview on 17 November 2026, where Google's launch blog says VPC Service Controls protect them and its security controls page says preview models get no security controls.
Section 09Agents, tools and MCP
Agents are where the design is tested hardest, because an agent holds access and makes its own choices about using it. Five rules address how an agent gets access and uses tools, and each has a counterpart in the checklist.
- One identity per agent. Agents that share an identity cannot be given separate permissions or be disabled separately. The UK National Cyber Security Centre's advice of 20 August 2026 asks for task-only permissions and short-lived credentials.
- The person's identity travels with the request. The agent swaps the user's token for one scoped to a single downstream service that names both user and agent (the on-behalf-of flow, or token exchange under RFC 8693).
- Tools are allowed by list, outside the agent. The list lives in a gateway or policy engine that the agent's prompt cannot change.
- Actions that cannot be undone wait for a person. Check where the approval is enforced, because Microsoft's toolbox page says its MCP endpoint passes approval flags along and leaves enforcement to the software that runs the agent.
- Credentials stay outside the agent's sandbox, the isolated environment it runs in. Anthropic's deployment guidance says to inject them through a proxy outside that boundary, which keeps reusable credentials out of a hijacked agent's reach.
MCP has its own authorisation rules, in a revision dated 28 July 2026 and built on the OAuth 2.1 standard. A server reached over HTTP is a protected resource. It must publish metadata that names its authorisation server, and it must check that each token was issued for that server and no other. A client must name the server it wants a token for, must check which authorisation server answered, and must use PKCE, a proof that the party redeeming a login code is the one that asked for it.
Two details of the specification matter for an audit. Passing a caller's token through to another service is banned. Authorisation is optional in the specification, so a team has to confirm that each HTTP server it uses has it turned on, while servers that run locally are expected to take credentials from their environment.
The week's disclosures show why. The AWS bulletin of 16 November 2026 describes an MCP server that did not restrict which endpoints it would connect to, so injected text could point it at an attacker's address. Research by Apiiro, reported by The Hacker News on 7 November 2026, found that when a user of Claude Code approves a command with the option to stop asking, the command is saved to a settings file with any password typed into it, and that 1,122 of 104,185 public copies held at least one secret. In both cases the control that helps sits outside the tool: a list of permitted outbound destinations in the first, and scanning repositories for secrets in the second.
Section 10The checklist: 40 checks and the evidence for each
The checklist has 40 checks in eight tables. Seven tables follow the control points and say where each check is set on Microsoft Azure, AWS and Google Cloud, and the eighth covers companies that call Anthropic or OpenAI directly. Each check can be answered yes or no and names the evidence that proves a yes, and the cloud cells state what each vendor's documentation said on 17 November 2026. A check passes only when someone other than the builder has seen that evidence.
| Check | Evidence | Microsoft Azure | AWS | Google Cloud |
|---|---|---|---|---|
| 1. Every agent in production has its own identity. | Agent inventory matched to the identity directory. | Entra Agent ID. Legacy Foundry agents share one identity per project. | AgentCore Identity, one workload identity per agent. | Agent Identity. On Agent Runtime it is opt-in, or the agent runs as a service account. |
| 2. No application or agent calls a model with a long-lived API key. | The deny policy in force and a key inventory showing none. | The disableLocalAuth setting plus the built-in Azure Policy that denies key access. Old keys can work for hours. | IAM deny on CallWithBearerToken in both the bedrock and bedrock-mantle namespaces, and on iam:CreateServiceSpecificCredential. | Constraint iam.managed.disableServiceAccountApiKeyCreation, on by default. Review its default exception for the Gemini API. |
| 3. Agent and workload credentials expire within a day and renew without a person. | Token lifetime settings and one decoded token showing its expiry. | Entra access tokens last 60 to 90 minutes by default. | STS sessions default to 1 hour. Short-term Bedrock keys last up to 12 hours. | Agent Identity certificates last 24 hours and renew themselves. |
| 4. Every key, role and app consent held by a vendor has a named owner and an end date, and unused ones are removed. | Vendor access register with the last sign-in date per entry. | Entra service principal sign-ins and app consent grants, both named in the FBI notice. | IAM role trust policies that name another account. | Workload Identity Federation in place of service account keys. A constraint disables keys found exposed in public. |
| 5. When an agent acts for a person, a token exchange carries that person's identity. | A downstream token or log entry naming both person and agent. | Entra on-behalf-of flow. In Foundry, an agent acting for the user is preview. | AgentCore Identity token exchange. Deny GetWorkloadAccessTokenForUserId in production. | Agent Identity auth manager, generally available since 22 August 2026. |
| 6. Every agent is in an inventory with a named human owner. | Registry export with an owner on every row. | Sponsors and owners on Entra agent identities. Lifecycle policies that disable inactive agents are preview. | AWS Agent Registry, with detection across AWS Organizations. | Agent Registry. |
| Check | Evidence | Microsoft Azure | AWS | Google Cloud |
|---|---|---|---|---|
| 7. Access policy is evaluated on each token or call, against the agent or the person it acts for. | A log entry showing the policy applied to an agent's request. | Conditional Access for agents, which needs an Agent 365 licence. Microsoft's pages disagree on whether it is preview. | IAM condition keys on Bedrock calls and JWT claim checks on AgentCore requests. | IAM Conditions and deny policies. Agent tokens fail outside their runtime. |
| 8. Each agent's permissions are tied to its task, with no project-wide or full-access role. | Role assignments per agent, each with its reason. | Azure role assignments on the agent identity. New Foundry identities start with none. | One execution role per runtime, with source account and source ARN conditions in its trust policy. | Avoid project-level aiplatform.user. Unified Access Policy grants are not limited by custom policies that restrict role granting. |
| 9. Only approved models can be deployed or called in any account. | A refused test deployment and a refused test call. | The built-in Azure Policy for approved models, in Deny mode. | Service control policy deny on bedrock:InvokeModel. Denying the Marketplace subscription alone does not block access. | Organisation policy vertexai.allowedModels, which covers Model Garden only. |
| 10. A stop test has shown how long it takes to disable one agent or revoke one credential. | Dated test record with the measured time. | Disabling a blueprint stops its agents. Key cutoff can take several hours. | AWS says short-term keys cannot be revoked one by one. Deny by token issue time, or deny the identity. | IAM deny policies, which are checked before any allow policy. |
| Check | Evidence | Microsoft Azure | AWS | Google Cloud |
|---|---|---|---|---|
| 11. The model endpoint accepts traffic only over a private route. | Network settings and a refused call from the internet. | Public network access Disabled, with a Private Link endpoint. | PrivateLink endpoints with a custom endpoint policy. Test that the public route is refused. | A VPC Service Controls perimeter. Private Service Connect alone leaves the public route open. |
| 12. Agents run in a network with no route to the internet unless one is added on purpose. | A blocked outbound test. | Agent Service private networking. Egress controls for hosted agents are preview. | AgentCore VPC mode, with no internet unless you add a NAT gateway. | A VPC Service Controls perimeter. Agents deployed before it are not secured. |
| 13. Every route that still leaves the private network is listed and approved. | Egress register with an approver per route. | Bing grounding, web search and SharePoint grounding use public endpoints. | Gateway calls to external MCP servers, and the Browser tool. | Inside a perimeter, internet access goes through your own proxy. |
| 14. For each third-party model, the company knows whether the cloud or the model maker hosts it. | Model register with the hosting option per model. | Claude hosted on Azure, or on Anthropic infrastructure. | Claude on Amazon Bedrock, or Claude Platform on AWS. | Claude run by Google, or by Anthropic at claude.googleapis.com. |
| Check | Evidence | Microsoft Azure | AWS | Google Cloud |
|---|---|---|---|---|
| 15. Agents can call only tools and MCP servers on an approved list enforced outside the agent. | The list and a refused call to an unlisted tool. | Azure Policy for approved and denied tools (preview). | AgentCore Gateway with Policy in AgentCore. | Agent Gateway with Agent Registry. Google's pages disagree on the status of per-tool rules. |
| 16. Tool policy denies by default, and tool traffic that skips the gateway is listed. | Gateway policy export and the list of exceptions. | The AI gateway (preview) routes only new MCP tools made in the portal without managed OAuth. | Gateway interceptors with default deny. | No wildcard rule for unregistered endpoints. |
| 17. An action that cannot be undone waits for a person, and the platform enforces the wait. | Approval records and a refused test. | The require_approval setting defaults to always for MCP tools. Toolbox flags are enforced only by the agent runtime. | Policy in AgentCore can refuse a tool call by its inputs before it runs. Show the approval step separately. | Rules can match a tool's destructive hint (preview). |
| 18. Each MCP server accepts only tokens issued for it and never forwards a caller's token. | Server settings and a refused token with the wrong audience. | With managed OAuth, Agent Service will not send Microsoft-audience tokens to third-party MCP servers. | Token passthrough is offered for gateway targets, so check which option is set. | Agent tokens are bound to their runtime, so a forwarded token fails elsewhere. |
| 19. Every MCP server in use has an owner and runs a version with all published security fixes. | MCP server inventory with owners and versions. | Microsoft says it does not test third-party MCP servers. | AWS published four MCP security bulletins from 11 to 16 November 2026. | Agent Registry holds the approved MCP servers. |
| 20. Calls to high-impact tools and APIs are rate limited. | A throttled test. | API Management rate limits through the AI gateway (preview). | AgentCore Gateway rate limits. | Apigee quotas per tool. |
| 21. Agents that run generated code use an isolated session with no long-lived credential inside. | Runtime settings and a search of the session for stored secrets. | Hosted-agent sandboxes hold no blueprint credential. | AgentCore Runtime gives each session its own microVM. | Agent Identity has no long-lived keys. Threat detection for container escapes is preview. |
| Check | Evidence | Microsoft Azure | AWS | Google Cloud |
|---|---|---|---|---|
| 22. Prompts and outputs pass an inspection service that can block, on every model endpoint in use. | A blocked test prompt per endpoint. | Prompt Shields and Foundry guardrails, for Models sold by Azure. Agent guardrails are preview. | Bedrock Guardrails, enforceable across accounts. They do not run on bedrock-mantle. | Model Armor, with an organisation policy that can force required filters. |
| 23. Text arriving through documents and tool results is inspected, or the gap is written down. | Inspection settings, or the signed exception. | Document detection is off by default. Tool response scanning is preview. | The prompt attack filter skips tool results. ApplyGuardrail can check any text. | The built-in integration does not screen documents. Use Model Armor on Agent Gateway. |
| 24. The team has tested what happens when inspection is skipped or unavailable. | The test record. | An agent's guardrail (preview) replaces the model's, so test the agent's. | Callers can skip the prompt check with input tags. | The built-in integration continues unscreened when Model Armor is unreachable. |
| 25. Sensitive data patterns are blocked or masked in prompts and responses. | A blocked test containing one. | Purview data loss prevention blocks prompts by sensitive information type. | Guardrails sensitive information filters. | Sensitive Data Protection inside Model Armor. |
| Check | Evidence | Microsoft Azure | AWS | Google Cloud |
|---|---|---|---|---|
| 26. Retrieval returns only documents the asking person may open. | A test with two users whose access differs. | Agents on Azure AI Search can honour sensitivity labels. | Access list filtering on knowledge bases. The application must authenticate the user. | Agent Search access control (preview). |
| 27. For each model, the provider's retention of prompts and outputs is known in writing. | Retention register per model. | Abuse monitoring states no period in days. Customers can apply for modified abuse monitoring. | Data retention mode, enforceable by service control policy. Claude Fable 5 models require up to 30 days for most customers. | Flagged Gemini prompts can be kept up to 90 days. Some partner models log all traffic for up to 30 days. |
| 28. Stored AI data is encrypted with company-held keys where the service allows, and gaps are listed. | Key settings and the list of exclusions. | Customer-managed keys. A hosted agent's session folder is not covered. | KMS keys, including for the AgentCore token vault and Memory. | Customer keys by model. Cache and abuse logs are not covered. |
| 29. The geography where prompts are processed is restricted, and global routing is a recorded decision. | Deployment list by type. | Global deployments can use any region. DataZone deployments stay in their zone. | Global inference profiles can route to any commercial Region. | Data residency is listed per model. Global endpoints carry no residency guarantee. |
| Check | Evidence | Microsoft Azure | AWS | Google Cloud |
|---|---|---|---|---|
| 30. Every model call is logged with the caller's identity, in every account. | Logging settings and one sample entry for each endpoint in use. | Diagnostic settings, which Azure Policy can deploy. | Model invocation logging, off by default, on bedrock-runtime only. | Data Access audit logs, off by default. |
| 31. Routes that leave no usable log are blocked or covered another way. | List of such routes and the control on each. | API keys carry no per-request user attribution. | Inference on bedrock-mantle is an opt-in CloudTrail data event. | Express-mode API keys produce no Data Access logs the customer can see. |
| 32. Each agent run leaves a trace of its steps, including every tool call. | One complete trace. | Agent tracing, preview for workflow and external agents. The AI gateway does not log tool traces. | AgentCore Observability to CloudWatch. | Agent Observability tracing, on by default for new ADK agents. |
| 33. Threat detection for AI workloads is on, and its alerts reach the security team. | One routed test alert. | Defender for Cloud AI threat protection, text only. | GuardDuty AI Protection. Its prompt injection finding needs Bedrock Guardrails. | Security Command Center AI Protection, Premium tier. Agent workload support is preview. |
| 34. An alert fires when someone removes a guardrail, turns off logging or adds a credential to an agent. | A test of each case. | Watch credential additions and app-only activity in Entra, as the FBI notice advises. | The GuardDuty Foundational plan flags guardrail removal and disabled invocation logging. | Alert rules built on Admin Activity audit logs, which are always on. |
| Check | Evidence | Anthropic | OpenAI |
|---|---|---|---|
| 35. Staff reach the model maker's apps and console only through company single sign-on. | A refused password login. | Switches require single sign-on for the Claude apps and for the Console. | Single sign-on in the Admin Console. A proxy header can limit staff to approved workspaces. |
| 36. Workloads use short-lived federated tokens, and static keys have a maximum lifetime. | Federation rules and the key policy. | Workload Identity Federation. An organisation policy caps key expiry. | Workload identity federation. Admins can cap the lifetime of new keys. |
| 37. API calls are accepted only from the company's network addresses. | A refused call from elsewhere. | IP allowlist, in beta. It does not apply on cloud routes. | IP allowlist for all API customers. Mutual TLS is also offered. |
| 38. The retention terms of every API feature in use are known, and features outside zero retention are listed. | Feature list with retention against each. | 30 days as standard. Zero retention excludes features such as Files, Batch, the MCP connector and hosted Managed Agents. | Abuse logs up to 30 days. Zero retention excludes features such as Assistants, Agents, vector stores, Files and Batch. |
| 39. Admin and activity logs flow to company security monitoring. | A sample event seen in the company's monitoring tool. | Compliance API activity feed, kept 6 years. | Audit logs API for admin actions. Compliance Platform for Enterprise and Edu. |
| 40. Coding agents run under settings only an administrator can change. | A refused local override. | Claude Code managed settings. The sandbox can be forced on. | Codex managed configuration, with network access off by default in its sandbox. |
Forty checks cannot all be done at once. The order below puts inventory and logging first, because every later step needs to know what exists and whether a change worked. It then deals with the two things the week's reports pointed at: standing keys held by vendors, and tools that act without a check.
- Days 1 to 5: list agents, model API keys and vendor access, and remove what nobody uses (checks 4 and 6, and the key inventory for check 2).
- Days 6 to 10: turn on model call logging and agent traces in every account (checks 30 to 32).
- Days 11 to 15: turn off static keys and move callers to short-lived credentials (checks 1 to 3).
- Days 16 to 20: close public access to model endpoints, write the egress register and record where each model is hosted (checks 11 to 14).
- Days 21 to 25: put tool calls behind a gateway with an allow list and add approvals (checks 15 to 17).
- Days 26 to 30: attach content inspection, turn on threat detection and run the stop test (checks 22 to 24, 33 and 10).
The remaining checks follow in the second month. Keep the evidence for each check in one place with the date it was collected, so that it can be shown to an auditor as it stands.
Section 11Where zero trust does not help, and the limits of what vendors document
Zero trust controls who can do what. Keeping a service running when a provider has a fault is separate work with its own plan, and this paper does not cover it.
Prompt injection has no complete fix, and the vendors say so. Anthropic's research page of 24 November 2025 states that "No browser agent is immune to prompt injection". OpenAI's guidance for agent builders says its mitigations greatly reduce the risk without fully removing it. The design therefore assumes that an agent can be turned, and relies on narrow access and a closed network to contain what a turned agent can do.
Logs show which identity did what, and they do not show intent. In the vendor role case, Wiz reported on 9 November 2026 that every request carried the correct external identifier and succeeded, so a successful sign-in alone would not tell the attacker from the vendor. Someone still has to define what unusual looks like, such as a vendor role used for the first time in months.
A further limit comes from outside the access path. Microsoft's research of 12 November 2026 showed in controlled experiments that 90 poisoned examples, about 1% of a fine-tuning set, made coding agents take unsafe tool actions in about 78% of trials on familiar agent setups, so a model from an unverified source can itself be the untrusted part.
Previews change. Microsoft's tool policies for Foundry agents and Google's semantic governance policies are preview, and Anthropic's IP allowlist is in beta. Vendor pages also disagree in the places we have marked, and we have not picked a side. This paper draws on vendor documentation and public reports read on 17 November 2026. We did not test the products for it.
Section 12How the frameworks map to the control points
The table maps each control point to the texts a risk or audit team is likely to cite. The mapping is our own reading, which applies their existing identity and resource controls to agents.
| Control point | NIST SP 800-207 tenet | CISA pillar or capability | OWASP risk | EU text |
|---|---|---|---|---|
| 1. Who is calling | Tenet 6: authentication and authorisation enforced before access, in a continuous cycle. Tenet 4 allows a service identity as the caller. | Identity | ASI03 Identity and Privilege Abuse. NHI7 Long-Lived Secrets. | Delegated Regulation (EU) 2024/1774, Article 20: a unique identity per person and system. |
| 2. What it may reach | Tenet 3: access per session with least privilege. | Identity | LLM06 Excessive Agency. | DORA Article 9(4)(c). Delegated Regulation 2024/1774, Article 21: least privilege and yearly access reviews. |
| 3. The path | Tenet 2: all communication secured regardless of network location. | Networks | ASI07 Insecure Inter-Agent Communication. | NIS2 Article 21(2)(j): secured communications. |
| 4. The tool | Tenet 1: every computing service is a resource. | Applications and Workloads | ASI02 Tool Misuse and Exploitation. ASI04 Agentic Supply Chain Vulnerabilities. | NIS2 Article 21(2)(d): supply chain security. |
| 5. The content in and out | Tenet 5: trust no asset by default and watch its integrity. | Applications and Workloads | LLM01 Prompt Injection. ASI01 Agent Goal Hijack. | AI Act Article 15(5): resilience against attempts to alter use, outputs or performance. |
| 6. The data | Tenet 1: every data source is a resource. | Data | LLM02 Sensitive Information Disclosure. LLM08 Vector and Embedding Weaknesses. | AI Act Article 15(5): measures against data poisoning and confidentiality attacks. |
| 7. The record | Tenet 7: collect information and use it to improve security. | Visibility and Analytics (cross-cutting capability) | ASI08 Cascading Failures. LLM10 Unbounded Consumption. | Delegated Regulation 2024/1774, Article 21: access reviews at least yearly, which this record supports. |
NIST's work on agent security is still at an early stage. Its concept paper on agent identity closed for comment on 2 April 2026.
On the EU side, NIS2 names zero trust in recital 89 as an example of cyber hygiene, and its implementing regulation of 17 October 2024 lists zero-trust architectures among secure development principles. The cybersecurity duty for high-risk AI systems in Article 15 of the AI Act now applies from 2 December 2027 for Annex III systems and from 2 August 2028 for Annex I systems, after the Digital Omnibus on AI moved the dates. This section is general information and not legal advice.
QueryNow helps companies put this design in place on Microsoft Azure and AWS, starting with the inventories that the checklist asks for first. Every engagement starts with one scoped workflow and a written definition of done.
What to do with this on Monday morning
- 1. List every agent, every model API key and every key or role held by a vendor, and put a named owner against each.
- 2. Remove the vendor roles and keys that nobody has used, and set an end date on the rest.
- 3. Turn on full logging of model calls in every account and Region before changing anything else.
- 4. Turn off static API keys for model calls and move callers to short-lived credentials.
- 5. Close public network access to model endpoints and write down every route that still leaves the private network.
- 6. Put tool calls behind a gateway with an allow list, and require a person's approval for actions that cannot be undone.
- 7. Attach prompt and output inspection to every model endpoint and test what happens when it is unavailable.
- 8. Run a stop test on one agent and record how long revocation takes.
Designed PDF
Download the PDF
The full paper is free to read on this page. The designed PDF arrives in your inbox, attached, within a minute.
Sources
- Microsoft Security Blog, High volume phishing campaign uses Microsoft 365 outage lures to target multiple industry sectors (16 November 2026)
- Microsoft Security Response Center, CVE-2026-124194 Microsoft Outlook Spoofing Vulnerability (10 November 2026)
- Microsoft Security Response Center, security update release data for November 2026, including CVE-2026-109492 Microsoft MCP Gateway (10 November 2026)
- Wiz, Trusted by Default: Qubole Cross-Account Access Abused in the Wild (9 November 2026)
- SecurityWeek, Over a Dozen Organizations Hit in Qubole Supply Chain Attack (10 November 2026)
- Cursor, An update on our recent security incident (13 November 2026)
- Barracuda, Forty-four hours: Anatomy of a machine-speed intrusion (11 November 2026)
- FBI, FLASH-20261113-001, Threat Overview of ARTEX Autonomous Penetration Testing Framework (13 November 2026)
- AWS Security Bulletin 2026-185-AWS, CVE-2026-126681 (16 November 2026)
- AWS, Security Bulletins list, including 2026-179-AWS, 2026-180-AWS, 2026-184-AWS and 2026-185-AWS (11 to 16 November 2026)
- The Hacker News, Claude Code's 'Don't Ask Again' File Exposes Passwords and Keys in 1,122 Public Copies on GitHub (7 November 2026)
- Microsoft Security Blog, Trained to betray: How post-training supply chain attacks hijack coding agents (12 November 2026)
- Microsoft, Ignite 2026 live news page (start time 17 November 2026)
- NIST, SP 800-207, Zero Trust Architecture (final, 11 August 2020)
- CISA, Zero Trust Maturity Model, Version 2.0 (cover dated April 2023, the revision table in the PDF says March 2022)
- NIST NCCoE, Accelerating the Adoption of Software and Artificial Intelligence Agent Identity and Authorization, concept paper (initial public draft, 5 February 2026)
- OWASP GenAI Security Project, OWASP Top 10 for LLM Applications 2025 (17 November 2024)
- OWASP GenAI Security Project, OWASP Top 10 for Agentic Applications for 2026 (9 December 2025)
- OWASP Foundation, Non-Human Identities Top 10, 2025 (no release date on page, opened 17 November 2026)
- Model Context Protocol, Specification, Authorization, revision 2026-07-28 (28 July 2026)
- NCSC (UK), Managing the cyber risk of agentic AI, blog (20 August 2026)
- European Union, Directive (EU) 2022/2555 (NIS2), Article 21 and recital 89 (Official Journal text, read 17 November 2026)
- European Commission, Implementing Regulation (EU) 2024/2690 (17 October 2024)
- European Union, Regulation (EU) 2022/2554 (DORA), Article 9 (Official Journal text, read 17 November 2026)
- European Commission, Delegated Regulation (EU) 2024/1774, Articles 20 and 21 (Official Journal text, read 17 November 2026)
- European Union, Regulation (EU) 2024/1689 (AI Act), Article 15 (Official Journal text, read 17 November 2026)
- European Commission, Regulatory framework for AI, including the Digital Omnibus on AI, Regulation (EU) 2026/1744 (page updated 19 October 2026)
- Microsoft Learn, Microsoft Entra releases and announcements (page dated 26 June 2026)
- Microsoft Learn, What's new in Microsoft Entra Agent ID (23 October 2026)
- Microsoft Learn, What are agent identities? (5 November 2026)
- Microsoft Learn, Agent identity blueprints in Microsoft Entra Agent ID (6 November 2026)
- Microsoft Learn, Agent OAuth flows: On-behalf-of flow (3 November 2026)
- Microsoft Learn, Agent identity concepts in Microsoft Foundry (28 October 2026)
- Microsoft Learn, Authentication and authorization in Microsoft Foundry (5 August 2026)
- Microsoft Learn, Access tokens in the Microsoft identity platform (14 May 2025)
- Microsoft Learn, Disable local authentication in Foundry Tools (31 May 2026)
- Microsoft Learn, Built-in policy definitions for Foundry Tools (page dated 2 October 2025, content refreshed 13 July 2026)
- Microsoft Learn, Conditional Access for Agents in Microsoft Entra (31 July 2026)
- Microsoft Learn, Configure Microsoft Entra agent identities for increased security (18 May 2026)
- Microsoft Tech Community, What's new in Microsoft Entra: November 2026 (6 November 2026)
- Microsoft Learn, How to configure network isolation for Microsoft Foundry (29 September 2026)
- Microsoft Learn, Set up private networking for Foundry Agent Service (15 June 2026)
- Microsoft Learn, What is a network security perimeter? (9 November 2026)
- Microsoft Learn, Guardrails and controls overview in Microsoft Foundry (20 October 2026)
- Microsoft Learn, Govern MCP Tools by Using an AI Gateway (19 August 2026)
- Microsoft Learn, Overview of MCP servers in Azure API Management (11 September 2026)
- Microsoft Learn, Set up MCP server authentication (4 November 2026)
- Microsoft Learn, Create and manage a toolbox in Microsoft Foundry (26 October 2026)
- Microsoft Learn, Connect agents to MCP server endpoints (28 October 2026)
- Microsoft Learn, Prompt Shields in Azure AI Content Safety (28 August 2026)
- Microsoft Learn, What is Azure AI Content Safety? (6 November 2026)
- Microsoft Learn, Use Microsoft Purview to manage data security and compliance for Microsoft Foundry (1 May 2026)
- Microsoft Learn, Customer-Managed Key Encryption in Microsoft Foundry (5 November 2026)
- Microsoft Learn, Data, privacy, and security for Foundry Models sold by Azure (18 May 2026)
- Microsoft Learn, Foundry Models sold by Azure abuse monitoring (13 May 2026)
- Microsoft Learn, Monitor Model Deployments in Microsoft Foundry Models (8 September 2026)
- Microsoft Learn, Agent tracing overview (28 August 2026)
- Microsoft Learn, AI threat protection in Microsoft Defender for Cloud (19 May 2026)
- Microsoft Learn, Foundry Models from partners and community (21 September 2026)
- Microsoft Learn, Data, privacy, and security for use of Anthropic Claude models in Microsoft Foundry (23 June 2026)
- Microsoft Learn, Deploy and use Claude models in Microsoft Foundry (21 September 2026)
- AWS documentation, Overview of Amazon Bedrock AgentCore Identity (no date on page, opened 17 November 2026)
- AWS documentation, Get workload access token (no date on page, opened 17 November 2026)
- AWS documentation, IAM Permissions for AgentCore Runtime (no date on page, opened 17 November 2026)
- AWS documentation, On-behalf-of token exchange with AgentCore Identity (no date on page, opened 17 November 2026)
- AWS documentation, Release notes for Amazon Bedrock AgentCore (entries through November 2026)
- AWS documentation, Configure inbound JWT authorizer (no date on page, opened 17 November 2026)
- AWS documentation, AssumeRole, STS API Reference (no date on page, opened 17 November 2026)
- AWS documentation, Amazon Bedrock API keys (no date on page, opened 17 November 2026)
- AWS documentation, Control permissions for generating and using Amazon Bedrock API keys (no date on page, opened 17 November 2026)
- AWS documentation, Handle compromised Amazon Bedrock API keys (no date on page, opened 17 November 2026)
- AWS documentation, Claude Fable 5.2 model card (no date on page, opened 17 November 2026)
- AWS documentation, GPT-6.1 Sol model card (no date on page, opened 17 November 2026)
- AWS documentation, Request access to models (no date on page, opened 17 November 2026)
- AWS documentation, Identity-based policy examples for Amazon Bedrock (no date on page, opened 17 November 2026)
- AWS documentation, Enforce the use of specific guardrails (no date on page, opened 17 November 2026)
- AWS documentation, Use interface VPC endpoints (AWS PrivateLink) for Amazon Bedrock (no date on page, opened 17 November 2026)
- AWS documentation, Use interface VPC endpoints for AgentCore (no date on page, opened 17 November 2026)
- AWS documentation, Configure Amazon Bedrock AgentCore Runtime and tools for VPC (no date on page, opened 17 November 2026)
- AWS documentation, Route model inference requests across AWS Regions (no date on page, opened 17 November 2026)
- AWS documentation, Amazon Bedrock AgentCore Gateway (no date on page, opened 17 November 2026)
- AWS documentation, Set up outbound authorization for your gateway (no date on page, opened 17 November 2026)
- AWS documentation, Policy in Amazon Bedrock AgentCore (no date on page, opened 17 November 2026)
- AWS What's New, Policy in Amazon Bedrock AgentCore generally available (3 March 2026)
- AWS documentation, Fine-grained access control for AgentCore Gateway (no date on page, opened 17 November 2026)
- AWS What's New, Next-generation AgentCore Runtime is the default (12 November 2026)
- AWS documentation, Detect and filter harmful content by using Amazon Bedrock Guardrails (no date on page, opened 17 November 2026)
- AWS documentation, Detect prompt attacks with Amazon Bedrock Guardrails (no date on page, opened 17 November 2026)
- AWS documentation, Use the ApplyGuardrail API (no date on page, opened 17 November 2026)
- AWS documentation, Inference endpoints supported by Amazon Bedrock (no date on page, opened 17 November 2026)
- AWS documentation, Apply cross-account safeguards with Amazon Bedrock Guardrails enforcements (no date on page, opened 17 November 2026)
- AWS documentation, Amazon Bedrock data retention (no date on page, opened 17 November 2026)
- AWS documentation, Amazon Bedrock abuse detection (no date on page, opened 17 November 2026)
- AWS documentation, Amazon Bedrock data encryption (no date on page, opened 17 November 2026)
- AWS documentation, ACL-aware retrieval on managed knowledge bases (no date on page, opened 17 November 2026)
- AWS documentation, Monitor model invocation using CloudWatch Logs and Amazon S3 (no date on page, opened 17 November 2026)
- AWS documentation, Monitor bedrock-mantle API calls using CloudTrail (no date on page, opened 17 November 2026)
- AWS documentation, AgentCore observability, audit logging, and tracing (no date on page, opened 17 November 2026)
- AWS documentation, GuardDuty AI Protection (no date on page, opened 17 November 2026)
- AWS What's New, Amazon GuardDuty AI Protection (14 July 2026)
- AWS Security Blog, Security Hub adds AI workload protection and multicloud support for Microsoft Azure (14 July 2026)
- AWS documentation, What is Claude Platform on AWS? (no date on page, opened 17 November 2026)
- Google Cloud documentation, Gemini Enterprise Agent Platform release notes (last updated 17 November 2026)
- Google Cloud documentation, Agent Identity overview (last updated 17 November 2026)
- Google Cloud documentation, IAM release notes (last updated 17 November 2026)
- Google Cloud documentation, Use Agent Identity with Agent Runtime (last updated 17 November 2026)
- Google Cloud documentation, Workload Identity Federation (last updated 17 November 2026)
- Google Cloud Blog, Helping you build and defend at machine-speed in the agentic era (4 November 2026)
- Google Cloud documentation, Manage API keys (last updated 17 November 2026)
- Google Cloud documentation, Organization policy constraints (last updated 10 November 2026)
- Google Cloud documentation, Get a Google Cloud API key (last updated 17 November 2026)
- Google Cloud documentation, Overview of IAM Conditions (last updated 17 November 2026)
- Google Cloud documentation, Deny policies (last updated 17 November 2026)
- Google Cloud documentation, Principal access boundary policies (last updated 17 November 2026)
- Google Cloud documentation, IAM Unified Access Policies overview (last updated 17 November 2026)
- Google Cloud documentation, VPC Service Controls with Gemini Enterprise Agent Platform (last updated 17 November 2026)
- Google Cloud documentation, Agent Gateway overview (last updated 17 November 2026)
- Google Cloud documentation, Route Agent Runtime traffic through Agent Gateway (last updated 17 November 2026)
- Google Cloud documentation, Semantic governance policies overview (last updated 17 November 2026)
- Google Cloud documentation, Manage MCP tool access with API products, Apigee (last updated 12 November 2026)
- Google Cloud documentation, Model Armor overview (last updated 17 November 2026)
- Google Cloud documentation, Integrate Model Armor with Gemini Enterprise Agent Platform (last updated 17 November 2026)
- Google Cloud documentation, Model Armor release notes (last updated 17 November 2026)
- Google Cloud documentation, Gemini Enterprise Agent Platform and zero data retention (last updated 17 November 2026)
- Google Cloud documentation, Abuse monitoring (last updated 17 November 2026)
- Google Cloud documentation, Security controls for Generative AI (last updated 17 November 2026)
- Google Cloud documentation, Data residency, Gemini Enterprise Agent Platform (last updated 17 November 2026)
- Google Cloud documentation, Set up data source access control (last updated 10 November 2026)
- Google Cloud documentation, Agent Platform audit logging information (last updated 17 November 2026)
- Google Cloud documentation, Log and share requests and responses (last updated 17 November 2026)
- Google Cloud documentation, AI Protection overview, Security Command Center (last updated 11 November 2026)
- Google Cloud documentation, Agent Platform Threat Detection overview (last updated 11 November 2026)
- Google Cloud documentation, Control access to Model Garden models (last updated 17 November 2026)
- Google Cloud documentation, Anthropic's Claude on Google Cloud models (last updated 17 November 2026)
- Google Cloud documentation, Use Claude models on Google Cloud (Anthropic managed) (last updated 17 November 2026)
- Google Cloud documentation, OpenAI GPT Models (last updated 17 November 2026)
- Google Cloud Blog, Google Cloud brings OpenAI models to customers (21 October 2026)
- Anthropic, Workspaces, Claude Platform documentation (no date on page, opened 17 November 2026)
- Anthropic, Authentication, Claude Platform documentation (no date on page, opened 17 November 2026)
- Anthropic, Workload Identity Federation, Claude Platform documentation (no date on page, opened 17 November 2026)
- Anthropic, WIF reference, Claude Platform documentation (no date on page, opened 17 November 2026)
- Anthropic, Set up single sign-on (SSO), support article (13 October 2026)
- Anthropic, Set up role-based permissions on Enterprise plans, support article (updated over a month before 17 November 2026)
- Anthropic, IP allowlisting, Claude Platform documentation (no date on page, opened 17 November 2026)
- Anthropic, Claude Platform on AWS, Claude Platform documentation (no date on page, opened 17 November 2026)
- Anthropic, Data residency, Claude Platform documentation (no date on page, opened 17 November 2026)
- Anthropic Privacy Center, How long do you store my organization's data? (1 July 2026)
- Anthropic, API and data retention, Claude Platform documentation (no date on page, opened 17 November 2026)
- Anthropic, Access audit logs, support article (15 June 2026)
- Anthropic, Compliance API, Claude Platform documentation (no date on page, opened 17 November 2026)
- Anthropic, Inference hooks, Claude Platform documentation (no date on page, opened 17 November 2026)
- Anthropic, Claude Platform release notes (no date on page, opened 17 November 2026)
- Anthropic, Deploy managed settings, Claude Code documentation (no date on page, opened 17 November 2026)
- Anthropic, Sandboxing, Claude Code documentation (no date on page, opened 17 November 2026)
- Anthropic, Data usage, Claude Code documentation (no date on page, opened 17 November 2026)
- Anthropic, Securely deploying AI agents, Agent SDK documentation (no date on page, opened 17 November 2026)
- Anthropic, Mitigating prompt injections in browser use, research (24 November 2025)
- Anthropic, Claude in Amazon Bedrock, Claude Platform documentation (no date on page, opened 17 November 2026)
- Anthropic, Claude Models on Google Cloud, Google managed, Claude Platform documentation (no date on page, opened 17 November 2026)
- Anthropic, Claude in Microsoft Foundry, Claude Platform documentation (no date on page, opened 17 November 2026)
- OpenAI, Single sign-on (SSO) setup for OpenAI products, help article (about 21 October 2026)
- OpenAI, Corporate network controls in ChatGPT Enterprise, help article (about 28 October 2026)
- OpenAI, Manage permissions in the OpenAI platform, developer documentation (no date on page, opened 17 November 2026)
- OpenAI, API keys and service accounts, developer documentation (no date on page, opened 17 November 2026)
- OpenAI, Production best practices, developer documentation (no date on page, opened 17 November 2026)
- OpenAI, Workload identity federation, developer documentation (no date on page, opened 17 November 2026)
- OpenAI, Mutual TLS, developer documentation (no date on page, opened 17 November 2026)
- OpenAI, OpenAI Mutual TLS Beta Program, help article (about 19 October 2026)
- OpenAI, IP allowlist, developer documentation (no date on page, opened 17 November 2026)
- OpenAI, Private Link, developer documentation (no date on page, opened 17 November 2026)
- OpenAI, Data residency, developer documentation (no date on page, opened 17 November 2026)
- OpenAI, Data controls in the OpenAI platform, developer documentation (no date on page, opened 17 November 2026)
- OpenAI, List audit logs, API reference (no date on page, opened 17 November 2026)
- OpenAI, OpenAI Compliance Platform for Enterprise and Edu customers, help article (about 3 November 2026)
- OpenAI, Admin controls, security, and compliance for plugins and apps, help article (about 5 November 2026)
- OpenAI, Lockdown Mode, help article (about 26 October 2026)
- OpenAI, Agent approvals and security, Codex documentation (no date on page, opened 17 November 2026)
- OpenAI, Managed configuration, Codex documentation (no date on page, opened 17 November 2026)
- OpenAI, Model, tool, and data controls with Terraform, developer documentation (no date on page, opened 17 November 2026)
- OpenAI, Guardrails and human review, Agents SDK documentation (no date on page, opened 17 November 2026)
- OpenAI, Safety in building agents, developer documentation (no date on page, opened 17 November 2026)
- OpenAI, Responses API support on Amazon Bedrock, help article (about 22 October 2026)