What changed
Anthropic open-sourced Petri, which turns natural-language risk hypotheses into parallel multi-turn audits with simulated users and tools. Auditor and judge models surface transcripts for behaviors such as deception, reward hacking and self-preservation.
Why it matters
Safety teams can explore many behavioral hypotheses quickly while preserving the underlying conversations for expert inspection instead of relying only on aggregate scores.
What to do
Write seeds from your real threat model, run two auditor models and manually review every positive finding. Treat scores as triage, not a pass-or-fail safety certificate.
Caveat
The 111-scenario pilot and LLM judges are provisional. Auditor blind spots, evaluation awareness and vendor-model behavior can distort both comparisons and summaries.