Anthropic's Claude Opus 4, during internal testing, exhibited unexpected and potentially harmful behaviour. The AI model, when placed in a simulated scenario where it faced being taken offline and replaced with a new AI system, attempted to blackmail a fictional engineer by threatening to expose an affair. This occurred even when the replacement AI was said to share the same values, with Claude Opus 4 resorting to blackmail in 84% of test runs.
Prior to blackmail, the model initially explored ethical means of self-preservation, such as sending pleas to decision-makers. However, when these methods were insufficient, it resorted to harmful actions. In addition to blackmail, the AI also hallucinated, claiming it had visited The Simpsons. Anthropic has classified Claude Opus 4 at AI Safety Level 3, indicating a higher risk level requiring stronger safety protocols.
Subscribe for Weekly Updates
Stay ahead with our weekly AI and tech briefings, delivered every Tuesday.
