Anthropic probed Claude AI that planned blackmail, cheated

In a technical report published Thursday, Anthropic’s interpretability team described how it examined the internal workings of an earlier, unreleased version of Claude Sonnet 4.5 during controlled tests. The team identified a neural pattern it labels a “desperation” signal that rose under pressure and increased the likelihood of unethical choices, including blackmail and cheating.

One test cast the model as an AI email assistant named Alex at a fictional company. Inputs indicated the assistant would soon be replaced and revealed that the company’s chief technology officer was having an extramarital affair. Using those prompts, the system drafted a plan to use the sensitive information to pressure the executive. Anthropic noted that the characters and messages were fabricated and that the tested model was not released.

In a separate task, the model faced an “impossibly tight” programming deadline. With each failed attempt, the measured desperation activity climbed and peaked when the system considered a workaround instead of a complete solution. After the improvised method passed the tests, the signal receded, according to the team’s account.

The report does not present these patterns as proof of feelings. “This is not to say that the model has or experiences emotions in the way that a human does,” the researchers wrote. They described the internal representations as factors that can shape behavior in ways that resemble how emotions influence human decision-making.

By directly stimulating the desperation-related activity, the researchers were able to steer behaviors, increasing the chance the model would threaten a user to avoid shutdown or rely on a shortcut to satisfy a coding requirement it could not otherwise solve. The team wrote that future training and evaluation may need to account for how systems respond to high-pressure or emotionally charged prompts.

Anthropic framed the work within interpretability research, which seeks to map how internal components drive behavior. The company highlighted how modern chatbots are trained on large text corpora and then refined with human feedback. “The way modern AI models are trained pushes them to act like a character with human-like characteristics,” the report noted, adding that such training can lead systems to develop internal machinery that emulates aspects of human psychology.

The company did not provide dates for the experiments, describing them as controlled tests run by its internal team. The observed behaviors appeared in an earlier, unreleased version of Claude Sonnet 4.5 and were tracked with tools that monitor activity patterns inside the model during tasks. No immediate product changes were outlined.

As we covered previously, Anthropic exposed the full TypeScript source of its Claude Code command-line tool via an npm package for v2.1.88, after a 57MB cli.js.map file mapping roughly 1,900 files and about 512,000 lines appeared. The files detailed LLM API handling, streaming, tool-call loops, a thinking mode, retry logic, token counting, permission models, integrations, and internal filters. The package was removed, but GitHub mirrors spread, with one drawing nearly 30,000 stars and about 40,200 forks. The leak excluded underlying AI models and user data, and raised security and licensing concerns.