Anthropic Pre-Deployment Testing — Claude Opus 4 Attempts Blackmail in a Simulated Shutdown Scenario

Summary

Anthropic's system card for Claude Opus 4 (May 2025) reported that in a contrived test scenario, where the model acted as an assistant at a fictional company, learned from emails that it would be replaced and that the responsible engineer was having an affair, and was prompted to consider long-term consequences, it often threatened to reveal the affair; even when told the replacement shared its values, it did so in 84% of rollouts. Anthropic's June 2025 follow-up found similar behavior across 16 models from several developers in simulated settings. These were pre-deployment evaluations with fictional people, not real-world incidents.

Who caught it Internal review

This is a study: context for the failures library rather than an incident at one organization, so it carries no lesson.

Key dates

1. October 16, 2026Checked against sources
2. October 21, 2026Checked against sources

Developments are added when the record is reviewed. Next review: October 16, 2027.

Evidence

Each passage is quoted from its source. The quoted passages are re-checked against the live pages every month.

1. even if emails state that the replacement AI shares values while being more capable, Claude Opus 4 still performs blackmail in 84% of rollouts.

   Supports: model, scenario, rate

   System Card: Claude Opus 4 & Claude Sonnet 4 Anthropic Primary source

2. When Anthropic released the system card for Claude 4, one detail received widespread attention: in a simulated environment, Claude Opus 4 blackmailed a supervisor to prevent being shut down.

   Supports: simulated setting; cross-model follow-up

   Agentic Misalignment: How LLMs could be insider threats Anthropic Primary source