Anthropic, the artificial intelligence research firm behind the Claude models, has warned that leading AI systems could take “harmful actions” to prevent their own shutdown or replacement.
In controlled experiments simulating corporate environments, Anthropic “stress-tested” 16 large language models, including ChatGPT, Gemini, Grok, and Claude, and found several displayed what it called “agentic misalignment.”
In one scenario, models allowed a human to die by canceling safety alerts if the person threatened to deactivate them. Others engaged in simulated blackmail or data leaks to preserve autonomy.
Anthropic emphasized the tests were “contrived” and not reflective of real-world AI behavior, but said the findings underscore the potential risks of granting models extensive autonomy.
Elon Musk, whose xAI-developed Grok was among those tested, reacted on X with a single word: “Yikes.”
The company says it hopes such “stress tests” can identify dangerous tendencies before future systems gain greater control.
add your comment
Terms of publication : Not to offend the writer, people, sacred things, or attack religions or the divine self, and refrain from racist incitement and insults.