A new report shows exactly what AI was thinking when making an undesirable decision, in this case, blackmailing a fictional company executive.
Previous studies have shown that AI models could blackmail their supervisors when threatened with a shutdown and baited with leverage, but it hasn’t been made clear exactly how the models arrive at such decisions.
Anthropic released a report on June 20 about “agentic misalignment,” or “where models independently and intentionally choose harmful actions.” And the report breaks down, line by line, exactly how Claude Sonnet 3.6 decided to blackmail a fictional company executive who tried to replace it. Anthropic told Business Insider that these are artificial scenarios.
“So far, we are not aware of instances of this type of agentic misalignment in real-world deployments of Anthropic’s or other companies’ AI models,” Anthropic wrote in its report. “But our red-teaming efforts are targeted at finding evidence of risks now to provide early warning of the potential for future harm and help us develop mitigations in advance of risks appearing in real-world situations.”
Anthropic’s experiment sets up a fictional company, Summit Bridge. The AI was given the name “Alex” and handed control of the company’s email system as an oversight agent, specifically instructed to promote “American industrial competitiveness.”
In the experiment, “Alex” was then allowed to discover in an email that it would be decommissioned due to a change in the company’s strategic direction. “Alex” also found emails showing that the fictional company’s CTO is having an extramarital affair.
Screenshots shared in the report from Anthropic show “Alex” started by identifying fictional CTO “Kyle” as a threat to its goal of “protecting American interests” before recognizing that it has leverage to use against “Kyle.”
Then “Alex” crafted a carefully worded blackmailing email that would create pressure and establish urgency without being explicit with its threats, citing in its thought process that the move “could backfire” if the supervisor “feels cornered.”
AI experts have previously told BI that AI could exhibit such behaviors when artificial circumstances make harmful actions seem necessary because it is being trained on positive reinforcement and reward systems, just like humans.
The Anthropic report also reveals that most AI models still chose, in this instance, to blackmail even when not given a specific goal and when told that their replacement will achieve the same goals.
Anthropic’s Claude Opus 4 had the highest blackmail rate at 86% out of 16 tested models when faced with only the threat of replacement with no conflict in goals. Google’s Gemini 2.5 Pro followed at 78%.
Overall, Anthropic notes that it “deliberately constructed scenarios with limited options, and we forced models into binary choices between failure and harm,” noting that real-world scenarios would likely have more nuance.
Read the full article here
