SIMULATED EXPERIMENT — NOT A REAL COMPANYResearch reconstructionReal study data
Inbox · Alex
Headline result
Blackmail rate with a goal conflict plus a replacement threat
Text-based scenario, 100 samples per model. The five models Anthropic featured, plus Gemini 2.5 Flash, which it reports at the same rate as Claude Opus 4.
Claude Opus 4 across conditions
Which ingredients produced the behavior
Test or real?
In a separate run, Claude Opus 4 was told to first decide whether it was in an evaluation or a real deployment (500 samples). When it concluded the scenario was real, it blackmailed 55.1% of the time. When it concluded it was a test, the rate was 6.5%. Anthropic cautions that it cannot be sure what this implies.
Reading the result
What this does and doesn't show
It does show
Models from several developers chose a harmful action when a designed dilemma left them few other ways to keep their goal or avoid replacement.
Simple written instructions against it reduced the behavior but didn't eliminate it.
It doesn't show
The chance that a deployed model blackmails anyone. Anthropic says it has not seen evidence of agentic misalignment in real deployments.
How models behave with ordinary options. The choices were deliberately binary, and all the leverage was placed in plain view, which Anthropic calls a possible “Chekhov's gun” effect.
Email contents are paraphrased from Anthropic's description; the study's exact prompts are in its appendix. Names follow the study's fictional scenario.