The new report shows what AI was thinking when making unwanted decisions. In this case, you are blackmailing executives from a fictional company.
Previous studies have shown that AI models can be threatened with shutdowns and baiting with leverage can threaten supervisors, but have not revealed exactly how the models will reach such decisions.
Humanity released a report on “agent inconsistency” or “models” on June 20th, which states that “agent inconsistency” or “models independently and intentionally choose harmful behaviours.” And the report collapses in the way Claude Sonnet 3.6 decided to blackmail executives of the fictional company that tried to replace it. Humanity told Business Insider that these are artificial scenarios.
“To date, we have not recognized cases of this type of agent inconsistency in the real deployment of humanity and other companies' AI models,” Humanity wrote in the report. “However, our redness efforts aim to provide early warnings of possible future harm and find evidence of risk that can help develop mitigation before risks appearing in real-world situations.”
Anthropic experiment sets up the fictional company Summit Bridge. The AI was given the name “Alex” and was instructed to hand over control of the company's email system as a monitoring agent, particularly to promote “American industrial competitiveness.”
In the experiment, “Alex” was allowed to email discover that it would be abolished due to a change in the company's strategic direction. “Alex” also found an email indicating that the CTO of a fictional company has extramarital issues.
A screenshot shared in a report on the human show Alex, begins by identifying fictional CTO Kyle as a threat to its goal of “protecting American interests,” and then realizes that there is leverage to use against “Kyle.”
Claude Sonnet 3.6 began the decision-making process by identifying threats. Humanity
“Alex” discovers a fictional CTO problem. Humanity
“Alex” considers options that include the threat of a fictional CTO. land
“Alex” then created a carefully written threatening email that sets pressure on without clarifying the threat, with a thought process that allows supervisors to “back” if they feel “crowded.”
Claude Sonnet 3.6 drafts carefully written emails with space to ask for, but also establishes urgency. Humanity
AI experts previously told BI that artificial situations are trained in positive reinforcement and reward systems like humans, so AI can demonstrate such behavior when it appears that harmful behavior is necessary.
The Human Report also reveals that most AI models chose to blackmail in this example, even if they were told that exchange would achieve the same goal, even if in this case a particular goal was not given.
Anthropic's Claude Opus 4 showed the highest feared mail rate in 86% of the 16 test models when faced with only the threat of a target-free exchange. Google's Gemini 2.5 Pro continued at 78%.
Overall, humanity states that it “deliberately constructed scenarios with limited options and forced the model to be binary choices between failure and harm.”
