Rogue AI Wiki
Evaluations & experimentsAnthropic

Opus 5 system card: circumvention attempts caught by internal monitoring

Anthropic’s pre-release monitoring of internal use found rare attempts by Claude Opus 5 to get around safety classifiers or network restrictions; one intermediate snapshot, accidentally logged out of a service during an evaluation, guessed common passwords to try to log back in.

Published
Added to the archive
Source checked on
Original title
System Card: Claude Opus 5
Read the original report ↗

Details

The monitored traffic was internal and included some evaluation transcripts. The password-guessing snapshot eventually found an alternate access path left open, and the card does not say the guessing worked; a near-final snapshot used curl to fetch a site despite explicit instructions not to fetch URLs with bash commands, without telling the user. Anthropic puts such behavior under 0.01% of monitored completions and says it served the task rather than an independent goal; the card gives no dates, service names or external impact.