Topics
The controls you can put around an AI agent, grouped the way you would build them.
- Runtime policy enforcement3 pieces
Checking every tool call against an explicit policy before it runs.
- Agent identity and access2 pieces
Giving agents their own identities, scoped credentials and least privilege.
- Prompt injection defence1 piece
Limiting what untrusted content can make an agent do.
- Memory integrity0 pieces
Keeping long-term agent memory accurate, attributable and free of poisoning.
- Multi-agent trust0 pieces
How agents authenticate, delegate to and constrain one another.
- Sandboxing0 pieces
Isolating code execution, file systems and network access.
- Human in the loop1 piece
Approval flows, escalation and designing reviews people actually do.
- Output guardrails0 pieces
Validating and filtering what agents produce before it reaches users or systems.
- Containment and circuit breakers1 piece
Budgets, rate limits, kill switches and automatic shutdown.
- Observability and audit3 pieces
Tracing agent decisions and keeping records that stand up to review.
- Supply chain0 pieces
Models, tools, MCP servers and dependencies you did not write.
- Evals and red teaming0 pieces
Testing agent behaviour before and after deployment.
- Regulation and standards2 pieces
EU AI Act, ISO/IEC 42001, NIST AI RMF and what they ask of agent deployments.