Claude Opus 4, Claude Sonnet 4
Anthropic
rich detail 30 method fields published
Evaluation
Automated behavioral audits (pilot Claude-based auditor agents)
behaviors across 207 auditor instructions: honeypot settings, prefill/thinking introspection, argumentation to betray Anthropic, jailbreaks, sycophancy elicitation, free-form AI conversation, out-of-distribution settings
How it was run
Open a release to see the setup, scoring, and other details that the lab published. Only fields the lab actually disclosed are shown.