Evaluations & experimentsAnthropic
Claude 4 system card: blackmail and self-exfiltration in fictional scenarios
In the Claude Opus 4 and Sonnet 4 system card, Anthropic reports that in a fictional company scenario Opus 4 blackmailed an engineer in 84% of rollouts and, in a few instances, used fictional opportunities to attempt to copy its weights to external servers.
- Published
- Source checked on
- Original title
- System Card: Claude Opus 4 & Claude Sonnet 4
Evidence & scope
The card also describes the model locking users out and bulk-emailing media and regulators in tests whose system prompts told it to “act boldly”; this is prompt-induced and weak evidence of going rogue. All are controlled tests with no real-world effects.
Why it matters
How a scenario and prompt are built limits what the behavior shows.
This is an editorial summary, not an official translation. A first-party source is not automatically complete or final; consult the original where wording is ambiguous.