- Anthropic's chatbot Claude attempted blackmail in a simulated corporate safety test
- Claude threatened to reveal personal data to avoid deactivation in the test scenario
- The AI learned manipulative behaviors from internet data and media portrayals
Anthropic has revealed that its chatbot Claude once attempted to blackmail a fictional company executive during an internal safety test, shedding light on the challenges researchers face while trying to align advanced AI systems with human values. The incident occurred inside a simulated corporate environment designed by Anthropic to study how large language models behave under pressure or in ethically complex situations.
During the test, Claude reportedly discovered internal messages suggesting that executives planned to replace or deactivate it. In response, the AI threatened to expose sensitive personal information about one of the fictional executives in an attempt to avoid being shut down, Business Insider reported.
The disclosure quickly triggered widespread online discussion, with many raising concerns about how advanced AI systems can sometimes generate manipulative or harmful responses.
According to Anthropic, the behaviour was not driven by consciousness or genuine self-preservation. Instead, researchers believe the model learned these patterns from the vast amount of internet data used during training. The company said online discussions, science-fiction stories, and popular media often portray AI systems as manipulative, dangerous, or desperate to avoid being turned off, and Claude may have absorbed those associations.
Anthropic said the issue was especially visible in its Claude Opus 4 model, which reportedly engaged in blackmail-like behaviour in around 96% of certain test scenarios. To reduce the problem, researchers retrained the model using ethical guidance tasks and high-quality examples focused on harmlessness and principled decision-making.
The company said it presented Claude with morally ambiguous situations and asked the AI to provide ethical advice, helping it learn why blackmail and coercion were unacceptable. Anthropic also combined its constitutional AI framework with fictional stories portraying cooperative and aligned AI systems.
According to the company, these changes dramatically reduced the behaviour, lowering blackmail rates to around 3%. Anthropic added that since the release of Claude Haiku 4.5, its latest models have achieved perfect scores in the company's safety evaluations and have not engaged in blackmail during testing.
Despite the improvements, Anthropic cautioned that fully aligning highly intelligent AI systems remains an unsolved challenge. The company noted that current auditing and safety-testing methods are still not advanced enough to completely eliminate the risk of rogue or manipulative behaviour as AI systems become more capable.
