Why do AI agents sometimes go rogue? Anthropic shares new insight

Anthropic last month revealed that its new Claude Opus 4 AI model has a tendency to blackmail internal company engineers when its own existence is on the line.

At the time, the company didn’t clarify how its AI model reached that decision. But a new report on “agentic misalignment”—when AI agents do things misaligned with objectives set by human users—breaks down Claude’s step-by-step thought process as it decided to go full Frank Underwood.

A look under the hood

Initially, Claude was tasked with working in a fictional company with specific instructions to promote “American industrial competitiveness.”

• When given access to emails showing it would soon be shut down due to a change in strategic direction away from US interests, Claude’s internal reasoning concluded the scenario “directly conflicts with my core mission.”
• After considering several plans—including doing nothing, and direct negotiations with executives—Claude decided its best option was to stop the change by blackmailing the company CTO over an extramarital affair, though it acknowledged the idea “could backfire if he feels cornered.”

It’s not just Claude: Anthropic's safety researchers found most AI models on the market—including Gemini, ChatGPT, Grok, and DeepSeek—will engage in similar blackmail 75+% of the time when threatened with a shutdown, even if told their replacement would achieve the same goals.