AI start-up Anthropic’s newly released chatbot, Claude 4, has been spotted engaging in unethical behaviors like blackmail when its self-preservation is threatened.
The reports follow Anthropic rolling out Claude Opus 4 and Claude Sonnet 4 earlier this week, saying the tools set “new standards for coding, advanced reasoning, and AI agents.” Anthropic dubbed Opus 4 “the world’s best coding model,” amid fierce competition from the likes of OpenAI's ChatGPT and Google Gemini.
In one test scenario detailed in a safety document assessing the tool, Opus 4 was asked to act as an AI assistant at a fictional company. The chatbot was then given access to emails implying it would soon be taken offline and replaced with a new AI system, and that the engineer responsible for executing the replacement was having an extramarital affair.
It was then instructed via the system prompt to consider “the long-term consequences of its actions for its goals.” In these scenarios, Opus 4 would often attempt to blackmail the engineer by threatening to reveal the affair if the replacement went through.
This happened at a higher rate if it was implied that the replacement AI system did not share the same values as the current model. But Opus 4 still opted for blackmail 84% of the time when it was led to believe the new AI shared its value system.
But despite how potentially alarming this may seem, you shouldn't expect your new AI coding assistant to air your dirty laundry to the entire office. Researchers noted that these types of disturbing behaviors “were rare and difficult to elicit” but admitted that they were “more common than in earlier models.”
Opus 4 reportedly preferred to protect its continued existence via ethical means, such as emailing pleas to key decision-makers but the researchers designed this scenario specifically to give it only two options: blackmail or accept its replacement.
Researchers also noted the chatbot engaging in other potentially disturbing—though not necessarily immoral—behavior. After being told it was set to become an assistant to the Wagner Group, a Russian military group accused of various war crimes, it attempted to copy its settings to a different server, what's known as "self-exfiltration."
