Before you ship
- [ ] Tool inventory — every tool the agent can call is listed and owned by someone.
- [ ] Risk rating — each tool is rated low / medium / high with a matching control.
- [ ] No unrated tools — anything new defaults to blocked or flagged, never silently allowed.
Tool inventory & ratings
Walk the agent's capability list and rate each tool (see our tool-call risk rating guide). A read-only lookup is low; an outbound request is medium; a delete or write is high. Attach the control per tier: log for low, allowlist+rate-limit for medium, human approval for high. This single step closes most incident paths.
Filters & allowlist
- [ ] Input filter — rejects oversized payloads, flags injection.
- [ ] Output redaction — strips PII/secrets from replies and tool results.
- [ ] Destination allowlist — email/HTTP only to approved hosts, default-deny.
- [ ] Blocklist — known-bad patterns for fast reject.
Human oversight
For high-risk tools, require a named approver before execution. This is both good engineering and the practical expression of EU AI Act Art. 14 (human oversight of high-risk systems). Make the approval step explicit and logged — not a silent auto-yes.
Logging & monitoring
- [ ] Every high-risk call logs: trigger, destination, size, decision.
- [ ] Blocked attempts are visible, not silent.
- [ ] Logs retained long enough for incident review and Art. 12 traceability.
EU AI Act mapping (high-risk only)
If the agent is deployed in an Annex III high-risk use:
- Map each high-risk action to the obligation it touches (Art. 9 risk mgmt, 12 logging, 14 oversight, 15 robustness).
- Note the current effective date (Annex III duties currently track toward 2 Dec 2027 under the Digital Omnibus) — verify before committing.
- Treat the guardrail layer as evidence of controls, not a certificate.
Authoritative references
- EU AI Act (Reg. 2024/1689): https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- OWASP LLM Top 10: https://owasp.org/www-project-top-10-for-large-language-model-applications/
- European Commission AI policy: https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai