Anthropic says its Claude Mythos model finds cybersecurity bugs so well that releasing it would be catastrophic. "The fallout – for economies, public safety and national security – could be severe," the company warned in an April blog post. But Heidy Khlaaf, chief AI scientist at the AI Now Institute, isn't buying it. She's spent her career building the exact code analysis tools Anthropic claims to have surpassed. The biggest red flag, she says, is the lack of false positive rates, a basic metric in security auditing.

The pattern is familiar. In 2019, when Anthropic CEO Dario Amodei was at OpenAI, the company refused to release GPT-2 citing safety concerns. They released it months later. OpenAI CEO Sam Altman later admitted those fears were "misplaced." Now Altman criticizes Anthropic's fear-based marketing while his own company follows the same playbook.

Shannon Vallor, professor of ethics of data and AI at the University of Edinburgh, sees the strategy clearly. "If you portray these technologies as somehow almost supernatural in their danger, it makes us feel like we are powerless," she told the BBC. "As if the only people we could possibly look to would be the companies themselves."

Emily M Bender, a computational linguistics professor at the University of Washington, calls it "unsubstantiated claims of power." The real harms get buried under the hype. Exploited workers. Environmental damage. Social systems we're actively breaking.

This fear narrative didn't come from nowhere. It's rooted in Effective Altruism philosophy, particularly the "longtermist" branch that worries about existential threats to humanity. Anthropic's founders were early EA supporters. OpenAI received major funding from EA-aligned donors like Dustin Moskovitz. The ideological framework, popularized by William MacAskill's "What We Owe the Future" and Nick Bostrom's "Superintelligence," frames AI risk in cosmic terms.

That framing has real consequences. When we're all worried about Skynet, we're not asking who's being watched. Who's being discriminated against. Who's cashing in.

For anyone tracking AI agents, the takeaway is simple. When companies tell you their technology is too dangerous to release, ask why they're telling you instead of just not releasing it. The danger narrative sells. It attracts talent, boosts valuations, and positions these companies as the only responsible adults in the room. The actual capabilities of these models matter less than the story around them.