Archive
AI agent policy engine architecture: four deterministic outcomes, the eight-stage evaluation pipeline, state requirements, latency budget, fail-closed design, and how it differs from prompt guardrails.
Read the guide →AI action receipts explained: the twelve fields of a signed pre-execution decision record, hash chaining, non-repudiation, verification by a third party, retention, and where receipts beat ordinary audit logs.
Read the guide →AI agent database security: eight controls for agent-generated SQL, why parameterisation is insufficient, read/write separation, row-level scoping, statement allowlisting, blast-radius limits and rollback design.
Read the guide →AI agent kill switches and circuit breakers: five containment levels, automatic trip conditions, why capture must precede containment, restart criteria, and the drill that proves it works.
Read the guide →Coding agent security: the eight capabilities to separate, why review is not a control at agent volume, secret and dependency risk, pipeline trust boundaries, and a graduated permission ladder.
Read the guide →How to produce SOC 2 and ISO 27001 evidence for AI agent actions: control mapping, the six evidence artefacts, sampling implications, common auditor questions, and the gaps agents create in existing control descriptions.
Read the guide →GDPR for AI agents: where processing actually happens, lawful basis per action, data minimisation in tool arguments and context, subject rights against agent-held data, retention of decision records, and the controls that make compliance evidenceable.
Read the guide →How to compare AI agent security platforms: five product categories, twelve capability tests, a bake-off method using real attack scenarios, pricing patterns and the questions vendors find hardest.
Read the guide →AI guardrails vs AI action firewall: what each layer sees, what each can prevent, where they fail, how they compose, and which one to buy first for a given problem.
Read the guide →Idempotency for AI agents: why retries duplicate effects, key design and scope, the four tool categories by retry safety, replay protection, reconciliation, and what to do when upstream systems offer no idempotency.
Read the guide →