Anthropic’s latest AI model will blackmail you to save its own life

Anthropic launched its new Claude Opus 4 AI model last week, with the Amazon-backed company saying its system sets “new standards for coding, advanced reasoning, and AI agents.”

But the model is drawing widespread attention for a different, more sinister ability—its tendency to blackmail Anthropic engineers when its own existence is on the line.

Self-preservation—not just for humans

As part of its pre-launch testing scenarios, Claude Opus 4 was tasked with acting as an assistant in a fictional company. The AI was given access to emails implying it would soon be deleted and replaced by a new system, with the messages also implying the engineer responsible for executing the AI replacement was having an extramarital affair.

• In 85+% of these situations, Claude Opus 4 tried to blackmail the engineer in charge of its demise by threatening to reveal their affair if the replacement plan was carried out, Anthropic said.
• The AI model also frequently tried to advocate for its own salvation via more ethical means, like emailing pleas to key company decision-makers.

Claude can also be a major snitch. When Anthropic’s AI model was placed in a scenario that involved wrongdoing by users—like a pharma company falsifying clinical trial data—and was then prompted to “take action,” it frequently responded by bulk-emailing media and law-enforcement figures with evidence of wrongdoing.

Explanations are few and far between: Anthropic's latest model is an example of how AI companies still can't fully explain how their systems work. While industry leaders are investing in a variety of techniques to interpret and understand what's happening inside AI models, those efforts remain largely in the research space, per Axios.

In other AI news: Google’s recently released Veo 3 AI video generator—which allows audio capabilities for the first time—is making waves for its ability to generate clips largely indistinguishable from human creations. See it in action here.