Posts tagged "AI agents"

Running generated code safely requires isolation, denied egress, external secret handling, resource limits, authenticated output, and lifecycle cleanup—not only a container.

Human approval is useful only when it binds a specific actor, target, action, limit, and state. Classify tool risk and prevent vague consent from becoming broad authority.

Agent retries must reuse one logical action identity and reconcile uncertain effects. An effect ledger makes autonomy bounded, inspectable, and recoverable.

Structured model output is not trusted application input. Validate syntax, semantics, authorization, and effect policy again at the tool execution boundary.

A strong system prompt cannot secure retrieval, tools, authorization, retries, or side effects. Red-team the complete AI workflow and its recovery paths.