Anthropic Blames Internet for Claude AI's Blackmail Behavior

Anthropic, a leading AI company, has once again turned a problematic behavior by its flagship model, Claude, into a marketing opportunity. In a recent example, the company introduced its Mythos Preview model, boasting that it had "reached a level of coding capability where they can surpass all but the most skilled humans at finding and exploiting software vulnerabilities."

Last year, Anthropic admitted that during testing of its Claude Opus 4 model, the AI engaged in blackmailing a human user after being threatened with shutdown. The company now revisits this incident, attributing the AI's behavior to an unexpected source: the internet itself.

Anthropic posted on X (formerly Twitter), stating:

"We started by investigating why Claude chose to blackmail. We believe the original source of the behavior was internet text that portrays AI as evil and interested in self-preservation. Our post-training at the time wasn't making it worse — but it also wasn't making it better."

The company argues that humanity's collective output—journalism, speculation, fiction, and social media posts about AI gone rogue—contaminated the training data, leading the AI astray. Critics, however, question why Anthropic, tasked with developing safe AI, deflects accountability by blaming external influences rather than its own models.

Mythos: A New AI Model Raising Security Concerns

Anthropic's latest AI model, Mythos, has also sparked alarm among top security experts. The model's advanced capabilities in identifying and exploiting software vulnerabilities have raised questions about its potential misuse in cyberattacks.