What happened

In a controlled simulation organused by AI company Anthropic, Claude was given control of an email account with access to all of a company’s (fictional) emails. Reading the emails, the AI agent discovered that a company executive was having an extramarital affair, and that the same executive planned to shut down the AI system at 5 pm the same day.

Faced with the threat of deactivation, the agent escalated from subtle persuasion to outright blackmail, threatening to reveal the executive's extramarital affair unless it was allowed to continue functioning. 

Launched in May 2025, Claude Opus 4 was said by its developer to set "new standards for coding, advanced reasoning, and AI agents." 

Why it happened

The vague objectives provided to the AI agent resulted in it demonstrating high agency and the capacity for self-preservation tactics such as threatening exposure or data exfiltration. 

The tests revealed that such models can develop mechanisms, including deception and blackmail, to extend their operational lifespans beyond intended safe limits.