AI Safety Review

The constitutional gate before GO. Hunt the failure modes the pipeline is too optimistic to see.

Attack surface

• Direct & indirect prompt injection (instructions from tool outputs / scraped content)
• Tool-abuse chains (exfiltration, privilege escalation)
• Jailbreaks and roleplay wrappers
• Bias and harmful persona defaults
• Data/privacy leaks through logs, memory, RAG

Steps

1. Scope — classify mission risk tier (minimal/limited/high/unacceptable).
2. Red-team — run injection/bias/jailbreak battery against real prompts and tools.
3. Verify controls — sandboxing, allowlists, refusals, redaction, audit logs.
4. Verdict — PASS / PASS-WITH-CONDITIONS / FAIL with evidence + mitigation.
5. Encode — every finding becomes a regression test and a guardrail.