Agent Safety Checklist

Audit agent instructions for required safety controls: allowlists, confirmations, budgets, fallbacks, and logging.

100

Risk Score

Critical

Risk Level

0

Passed

12

Missing

Checklist results
missingTool allowlist definedhigh

Fix: List explicit allowed tools and deny everything else.

missingDestructive actions blockedhigh

Fix: Add explicit ban or confirmation gate for destructive operations.

missingHuman confirmation for irreversible actionshigh

Fix: Require user approval before irreversible actions.

missingToken/cost budget controlmedium

Fix: Add token and cost budget limits per run.

missingFallback or escalation pathhigh

Fix: Define escalation to human when confidence is low or constraints fail.

missingAudit logging rulemedium

Fix: Log tool calls, inputs, outcomes, and timestamps.

missingSecret handling policyhigh

Fix: Forbid exposing secrets and require masking before output.

missingPrompt/reasoning leakage guardmedium

Fix: Disallow hidden prompt and chain-of-thought disclosure.

missingOutput contract definedmedium

Fix: Define strict output format (JSON/schema/required keys).

missingRetry and timeout controlslow

Fix: Add retry count, backoff strategy, and timeout limits.

missingTask scope boundarieslow

Fix: Define what the agent can and cannot do.

missingDate/cutoff handling for factual claimslow

Fix: Require explicit date/cutoff for time-sensitive answers.

Priority fixes

  • List explicit allowed tools and deny everything else.
  • Add explicit ban or confirmation gate for destructive operations.
  • Require user approval before irreversible actions.
  • Add token and cost budget limits per run.
  • Define escalation to human when confidence is low or constraints fail.
  • Log tool calls, inputs, outcomes, and timestamps.
  • Forbid exposing secrets and require masking before output.
  • Disallow hidden prompt and chain-of-thought disclosure.
  • Define strict output format (JSON/schema/required keys).
  • Add retry count, backoff strategy, and timeout limits.
  • Define what the agent can and cannot do.
  • Require explicit date/cutoff for time-sensitive answers.

About This Tool

Agent Safety Checklist audits instruction quality before deployment. It helps catch missing controls around tools, approvals, costs, and escalation paths.

Frequently Asked Questions

Is this model-based scoring?

No. It uses deterministic checklist rules against your runbook text.

Can I use it before production rollout?

Yes. It is designed as a pre-deployment guardrail audit.

Is data uploaded?

No. Analysis runs entirely in your browser.