AI agent security review guide
AI agents can read data, use tools and trigger workflows. Security design should examine what the system can actually do, not only what the model says. This guide turns common controls into a discovery checklist for business and technical teams.
Map what the agent can reach
Inventory user identities, data sources, memory or retrieval content, external APIs, tools, service accounts and write locations. For each connection, record an owner, data classification, purpose, allowed actions and approver. Isolate data by user or session.
Constrain tools with least privilege
Enable only the tools and resources needed for the task. Read access should not imply permission to modify or delete. Separate read, draft, write and external-send permissions. Remove unused tools and constrain APIs, file paths, recipients and operation counts.
Treat generated output as untrusted until checked
Documents, email, webpages and API responses may contain malicious or misleading content. Validate agent output against schemas and business rules before passing it to a database, code, payment, customer record or public channel. Model output must not become its own authorisation.
Require human approval for high-impact actions
Separate recommendations from execution for transfers, permission changes, data deletion, external messages, official-record changes and publication. Use explicit approval and traceable action logs. Set tool-call, retry, recursion, time and cost limits; test exceptions and human hand-off.
Review risk after material changes
Changes to models, prompts, retrieval data, tools, memory, permissions or third-party services can change risk. Record test cases, expected results, exception handling and approvers. After launch, review refusals, human corrections, unusual tool calls and cost. Measures help detect issues; they are not promised outcomes.
References
SRAA readiness guide
SRAA readiness guide · System security engineering · AI Agent workflow planning