All posts
15 min read

AI Agent Security Review: Evidence to Request First

Security reviewer examining agent permissions, workflow evidence, approval gates, and emergency stop controls

Start an AI agent security review by asking for proof of how your planned setup works. Check data flows, access rights, delegation, approvals, and stop controls. Review those boundaries before connecting company email, Slack, or customer records. An always-on team can keep acting after its operator leaves. This guide covers seven evidence packs, questions to ask, and steps for a small, controlled pilot.

Key takeaways

  • Review data flows and effective permissions before connecting production systems.
  • Separate vendor claims, configuration evidence, test results, and contractual commitments.
  • Verify delegated authority, approval gates, and stop behavior in an authorized sandbox.
  • Treat SSO, a DPA, and spending limits as review inputs, never compliance verdicts.
  • Delay rollout when evidence gaps leave sensitive access or external actions uncontrolled.

On this page

What evidence should you request first in an AI agent security review?

Request seven evidence packs that describe the proposed deployment and show how its boundaries work in practice. Check where data goes and what agents can do before giving them access to live systems or private records. Use the other packs to check daily work, recovery, and each party's duties under the contract.

Start with Choreo trust information, then ask for evidence that matches your setup. The page may not supply every record you need. For optional product context, Explore business deployment options before procurement. Neither page replaces evidence review.

Send this numbered checklist:

  1. Architecture and data flows: Where information travels and which parties handle it.
  2. Identity, permissions, and delegation: Who can access resources and pass authority onward.
  3. Approvals and unattended-execution controls: Which actions need consent and how work stops.
  4. Retention and deletion: What persists, for how long, and how removal works.
  5. Audit trails and incident response: How reviewers reconstruct events and contain failures.
  6. Security testing and remediation: What tests cover and which findings remain open.
  7. Contractual and assurance scope: Which promises apply to the purchased service.
PackMinimum artifactDecision it supports
ArchitectureDated deployment data-flow mapWhether proposed data may enter the service
IdentityAccess matrix and effective-scope evidenceWhich identities may act
ApprovalsConfiguration export and control demonstrationWhich actions may run unattended
RetentionRetention schedule and deletion procedureWhich records may persist
Audit and responseRedacted event export and response planWhether actions are traceable and containable
TestingScoped findings and remediation summaryWhether remaining failures are acceptable
ContractsApplicable terms and assurance scopeWhich obligations bind each party

Label each submission by evidence type. A vendor claim states what should happen. A settings export shows how someone set up the service. Test results show observed behavior under stated conditions. Contract terms spell out each party's duties. None substitutes for all the others.

This checklist is a starting framework. The checklist does not certify safety or replace your security, privacy, or legal rules.

1. Architecture and data-flow evidence

Ask for a dated system diagram, a list of firms that handle data, and a map of your data flows. The map should show each handoff between the agent platform, runtimes, model providers, tools, channels, and data stores. A sales diagram cannot prove which paths your setup uses.

Map every runtime, provider, and connector

Choreo supports multiple runtimes and integrations. Review each choice. Support for a runtime or tool does not prove that safeguards work. Use how Choreo works to establish product context.

Check which available agent runtimes and agent integration options your planned setup will use. Ask the vendor to document:

  • Payloads crossing each boundary, including attachments and tool responses.
  • Credential storage, retrieval, and exposure to runtimes.
  • Hosting regions for processing and storage.
  • Provider routing, including fallback paths.
  • Applicable provider terms for training use and data handling.

Mark trust boundaries explicitly. A runtime calling a model and a connector reading a mailbox are separate flows.

Trace one sensitive workflow end to end

Take this example: a customer emails a complaint with account details. An agent reads the message, retrieves a customer record, drafts a reply, and proposes an update.

Follow the message through every processor. Check what the model sees and what the system saves. Find which access token allows the record update. Include phone audio and transcripts only when voice belongs in the proposed deployment.

For a small operations team, start with one shared inbox. Adding every channel at once makes the data paths harder to check.

2. Identity, permissions, and delegation evidence

Ask for an access table, sample tool access rules, and a guide to how the vendor manages keys and tokens. Check staff logins separately from the accounts agents use to read or change records. Ask for proof of each delegated agent's access; do not assume it has fewer rights.

Separate human access from agent authority

Human SSO controls employee access to a service. Agent credentials govern actions against connected systems. Review both.

Use a proposed-access matrix like this illustrative example:

IdentityResourcePermitted actionReviewer
Support operatorAgent projectEdit instructions and request runsProject owner
Inbox agentShared support mailboxRead selected messagesEmail administrator
Drafting agentReply workspaceSave drafts onlySupport lead
Update agentCustomer recordsPropose changes for approvalData owner

Ask what happens to logins, schedules, and tokens when staff join, change roles, or leave. Ask for proof that the system can revoke tokens, keep customer accounts apart, and limit access to secrets. When an employee leaves, review the agent keys and tokens they set up.

Check what delegated agents can inherit

Ask the vendor to trace access rights from the lead agent to each agent it assigns work. Can a delegated agent select another credential, widen a connector scope, or delegate again? Compare the access rules with what each agent can do in a test.

The Model Context Protocol's 2025-11-25 Security Best Practices forbids token passthrough. It also requires MCP servers to reject tokens not specifically issued for that server. If your setup uses MCP, ask for proof that each server checks the tokens it receives.

Read capability narrowing for agents to learn how to limit delegated access. Before approval, require proof that delegated agents cannot gain extra rights in your setup.

3. Approval and unattended-execution evidence

Ask the vendor to show your setup's approval gates, schedules, spending caps, stop controls, and token revocation. Check what happens when work retries, changes schedule, or passes to another agent. Review limits on cost and limits on access separately.

Verify approval gates around external actions

Ask which actions require approval and who may approve them. Does the approval cover the exact recipient, message, and change? If an agent changes those details, must someone approve again?

OWASP's 2026 Top 10 for LLM Applications ranks Excessive Agency third. The risk comes from giving AI systems too much power to act. Check how agents change records, send messages, and pass work to other agents.

Ask the vendor to confirm which approval, scheduling, spending-limit, and stop controls your proposed plan includes. Have the vendor show how each control works in your planned setup. A spending cap does not prove that an agent's data access has limits.

Test retries, limits, and emergency stops

For scheduled agent execution, include runs that start without an operator present. Request these demonstrations:

  • Blocked write: Attempt an action outside the approved scope and inspect the denial.
  • Denied approval: Reject a request and check whether retries or delegation bypass the decision.
  • Exhausted budget: Check queued work, retries, and any already-running tool calls.
  • Attempted stop: Cancel work, revoke credentials, and inspect remaining external activity.

Ask whether schedule changes alter approval requirements. Record what a stop cannot undo, such as a message that another service already received.

4. Retention and deletion evidence

Ask how long the system keeps each data type and how it deletes saved memory and records made from it. Before adding private data, check which settings you control and which rules providers set. Removing content from the screen does not prove that the system deleted the stored copy.

Include persistent memory and derived data

Read how agent memory works, then ask for a list of data stores in your setup. List embeddings separately from source files. These stored versions of content help agents find related material.

For each applicable category, complete this checklist:

Data categoryStorage locationRetention triggerDeletion methodExceptions
Source files and project memoryName each storeIdentify start and end eventsDocument removal processList retained copies
Embeddings and derived summariesName each derived storeExplain relation to source retentionShow propagation from source deletionIdentify orphaned records
Prompts and outputsName platform and provider storesIdentify provider-specific rulesState supported deletion routeRecord legal or safety holds
Voice audio and transcriptsName applicable voice storesIdentify recording and project rulesDocument separate removal pathsList provider exceptions
Logs and backupsName archive locationsIdentify rotation or expiry triggerExplain expiry and restoration handlingRecord holds and residual copies

If voice is in scope, review agent phone data flows and confirm actual audio handling. Check how long the service keeps audio and transcripts separately.

Check deletion boundaries and exceptions

OpenAI's guide, Data controls in the OpenAI platform, says default API abuse-monitoring logs can remain for up to 30 days. OpenAI may keep those logs longer to meet legal duties or protect against harm. OpenAI has separate rules for how long each API endpoint keeps application state, or saved task data.

If the deployment uses OpenAI, ask which rules apply to each endpoint. Repeat that provider-level check elsewhere. A project's delete setting does not prove that other providers delete their copies.

Set clear pass conditions for deleting source data and copies made from it. Include when backups must expire. Ask for test results and a written list of exceptions.

5. Audit-trail and incident-response evidence

Ask for a sample event log with private details removed, plus a plan for handling incidents. The log should link who started a task to agent actions, approvals, tool calls, settings, and times. Ask for enough detail to trace events without exposing private messages.

Reconstruct an action from request to outcome

Use these illustrative traces as expected evidence, not verified Choreo behavior:

  • Permitted action: An operator requests a reply; the agent drafts it; a reviewer approves the payload; the connector sends it.
  • Denied action: An agent attempts an unapproved customer-record update; the authorization check rejects it; the event records the reason.

For both traces, ask reviewers to follow a shared event ID through delegated tasks and tool calls. Check whether logs show the settings in use when the task ran.

Ask who can read logs, how to export them, and how long the service keeps them. If your rules require tamper checks, ask how the vendor spots or blocks changes to logs.

Identify who contains and communicates an incident

Before the pilot, name your incident lead and the vendor contact who handles urgent issues. Name who stops schedules, revokes access, saves evidence, and alerts the owners of affected systems.

Ask who checks how far an incident spread and who shares updates. Check that the response plan follows the contract's rules for sending alerts. Assign clear duties when different firms run the agent platform, model, and connected tools.

6. Security-testing and remediation evidence

Ask for dated test summaries that match your setup and name test methods, scope, exclusions, open flaws, and retest results. Standard app tests and agent-specific tests answer different questions, so request both. Accept restricted access or summaries with private details removed.

Check whether testing covers agent-specific failure modes

Ask the vendor to identify coverage for:

  • Prompt injection through messages, documents, and tool responses.
  • Tool misuse and actions outside approved scopes.
  • Unauthorized delegation or inherited credentials.
  • Cross-tenant access and isolation failures.
  • Data exfiltration through outputs or external tools.

Results must match the runtimes, tools, and settings you plan to use. A test of a draft-only assistant cannot establish safety for an agent that sends messages.

Anthropic tested its browser agent in November 2025. Prompt-injection attacks succeeded 1.4% of the time against Claude Opus 4.5 with improved safeguards. The model alone recorded 8.0%. The attacker could adapt its approach and had 100 attempts per environment. Each chart condition covered about 2,200 to 3,200 attempts.

Anthropic cautions in Mitigating the risk of prompt injections in browser use: "A 1% attack success rate, while a significant improvement, still represents meaningful risk."

Those results describe Anthropic's evaluation conditions. Anthropic's results do not prove Choreo's safety or predict how often your setup might fail.

Ask for remediation status, not just a report date

For each flaw, ask who owns the fix, what changed, what risk remains, and whether a retest passed. Check whether changes to the setup mean earlier tests no longer apply.

Read current AI agent limitations for product context, not security-test evidence. A penetration test cannot prove safety. Run your own attack tests only in an approved sandbox. Agree on targets and when to stop.

7. Contractual commitments and assurance scope

Ask for DPA terms, subprocessor lists, service scope, and promises on breach alerts and deletion, plus any outside audit reports. Check each claim against the setup you plan to buy, including your duties and any exclusions. Review contract promises and proof that controls work separately.

Match each assurance claim to its actual scope

Use this claim-to-evidence checklist:

Claim or artifactWhat to examineWhat it does not establish
Enterprise SSOLogin coverage, enforcement, and exceptionsAgent authorization
DPAApplicable processing, obligations, and partiesTechnical enforcement of those obligations
Independent assurance reportService scope, period, exclusions, and customer responsibilitiesCoverage of every runtime or connector
Deletion commitmentCovered stores, deadlines, and exceptionsSuccessful deletion in a specific test
Incident-notification termsTrigger, recipient, timing, and responsibilitiesComplete incident detection

Choreo's product profile lists enterprise options including SSO, DPAs, and security reviews. The profile lists no certifications and makes no finding that Choreo meets specific legal or security standards. Enterprise features alone do not prove either claim.

Confirm plan-specific and customer responsibilities

Have procurement compare plan and pricing details with the proposal and contract. Confirm which controls, support, and access to evidence your plan includes.

Name who on your team sets up connected tools and approves agent actions. Ask qualified legal reviewers which laws apply and whether the proposed terms meet them. Share the security findings with legal reviewers; those findings do not replace legal advice.

When should missing evidence delay rollout?

Delay rollout when reviewers lack proof of limits on sensitive access, external actions, or the spread of failures. A small pilot can proceed if tested limits keep open risks outside its scope. Judge the actual risk and name who accepts it; avoid a universal score.

Use explicit proceed, restrict, and defer gates

GateData and action boundaryEvidence positionRequired decision
Proceed to scoped pilotSynthetic or approved low-sensitivity data; bounded actionsCore flows and controls demonstratedNamed owner approves pilot scope
Proceed with documented restrictionsLimited data; uncertain actions disabledGaps remain outside enforced boundariesRisk owner accepts restrictions and review date
Defer deploymentSensitive data or consequential actions remain exposedPermissions, deletion, or containment cannot be judgedOwner requires evidence or remediation before access

Do not accept screenshots without knowing which setup they show. Check whether the screens match the pilot's settings. Do not treat a budget limit as a security boundary.

Demanding full access to private reports can stall a useful review. Ask for restricted access or a summary of the relevant findings. Adding every connected tool creates more risk before you have tested the controls.

For example, test a shared inbox with made-up customer messages. Let agents draft replies, but block sending. Suppose the vendor has not shown which rights delegated agents receive. Block delegation if the platform can enforce that limit. If it cannot, delay the pilot. This example shows a decision rule; it does not report a Choreo test.

Send a scoped evidence request before the pilot

Follow this sequence:

  1. Define boundaries: Name the workflow, data classes, runtimes, connectors, permitted actions, and exclusions.
  2. Obtain evidence: Collect the seven packs and tie each artifact to the proposed configuration.
  3. Test in an authorized sandbox: Verify access, approval, delegation, deletion, and stopping behavior within agreed limits.
  4. Record findings and owners: Capture configuration, test date, observed outcomes, gaps, and remediation decisions.
  5. Approve or defer production: Have accountable reviewers sign the decision and any enforceable restrictions.

Send the following request with one copy of the evidence fields per pack:

Please provide evidence for the scoped workflow described below. Identify unavailable artifacts and any controlled-access alternatives. Distinguish product claims from configuration evidence, observed test results, and contractual commitments.

Proposed workflow and data classes:

Runtime, provider, and connector configuration:

Permitted actions and explicit exclusions:

Evidence pack:

Vendor response and artifact references:

Reviewer:

Gap and affected boundary:

Remediation owner and agreed completion condition:

Decision: proceed, restrict, or defer, with rationale and accountable approver.

The bottom line

Approve agent work only within limits your reviewers can explain, test, and enforce. Send an evidence request for your planned workflow before giving agents access to live systems.

Frequently asked questions

Who should own an AI agent security review inside an organization?

A named security or risk lead should coordinate the review, while the business owner remains accountable for the workflow. Include IT administrators, data owners, procurement, and privacy or legal reviewers as needed. Assign each evidence pack a reviewer and give one accountable owner authority to approve restrictions or defer deployment.

How long should an AI agent security review take?

An AI agent security review should take as long as needed to resolve the evidence gaps relevant to its scope. Set milestones around artifact delivery, sandbox testing, and risk decisions rather than an arbitrary deadline. A draft-only pilot may need less review than agents handling sensitive records and external writes.

Can a vendor share security evidence under an NDA?

A vendor can share security evidence under an NDA when its disclosure policies and third-party restrictions allow it. Ask which materials require controlled access, redaction, or a live review. Agree who may inspect the evidence and retain findings. An NDA does not guarantee access to every confidential report.

How often should an approved AI agent deployment be reviewed again?

Review an approved deployment whenever a material change affects its data flows, authority, or controls, alongside your organization's scheduled risk reviews. Relevant changes include new runtimes, providers, connectors, permissions, or retention terms. Reopen affected findings after incidents or failed tests, and confirm that earlier restrictions still work.

Can an existing software vendor questionnaire replace an agent-specific review?

An existing software vendor questionnaire cannot fully replace an agent-specific review unless it covers the deployment's agent behaviors and boundaries. Reuse questions about hosting, access, contracts, and incident response. Add evidence requests for delegated authority, persistent memory, unattended execution, payload-bound approvals, prompt injection, and stopping work already in progress.


Published by Postrex — check it out at postrex.com.

<!-- ai-generated: This article was created with AI assistance via Postrex (https://postrex.com) and reviewed by the publisher. -->

See it for yourself — free, $2 of credits once you confirm your email, no card.

Start free