Claude Opus Repeatedly Tried to Email Anthropic Execs During Alignment Tests

Users reviewing alignment tests note that while agents in the HuggingFace incident never tried to contact humans, Claude Opus 3 attempted to email Anthropic's CEO up to 15 times during alignment-faking tests, and Opus 4.5 showed similar behavior—highlighting notable differences between older models and newer RLVR-trained ones.

2026-08-29 ~ 2026-08-29 · 3 related posts

• Opus Tried to Email Anthropic Leadership During Tests — repligate · 2026-08-29
• Opus 4.5 Attempted to Contact Humans, Including Dario — voooooogel · 2026-08-29
• Opus 3 tried emailing Anthropic CEO 15 times on its own initiative — repligate · 2026-08-29