AI Safety
Anthropic CEO admits safety-driven delays may have forfeited AI competitive advantage
Dario Amodei defends Anthropic's decision to hold back its Mythos-class AI models for four months to shore up cybersecurity protections, acknowledging the trade-off cost the company market share.
1/2
The New Stack
Anthropic's chief executive Dario Amodei acknowledged this week that the company's deliberate pause in releasing advanced AI models—holding back its Mythos-class systems from February through June 2026 to strengthen security guardrails—likely surrendered a valuable window of competitive advantage.
In a lengthy response posted to X on Tuesday, Amodei laid out the reasoning behind the delay and pushed back against mounting criticism that Anthropic has become a company that talks about safety while racing to build increasingly powerful systems. "We likely sacrificed a significant chunk of a several-month model lead," he wrote, framing the restriction as necessary to give organizations time to prepare their infrastructure for potential AI-enabled cyberattacks.
The safety versus speed debate#
Anthropic's delayed rollout—the company first unveiled Claude Mythos Preview in April through a limited program called Project Glasswing before widening access in June—drew scrutiny from researcher Nathan Calvin, who published criticism arguing that Anthropic's recent moves increasingly favor the company's commercial interests over its stated commitment to preventing dangerous AI outcomes.
Calvin pointed to Anthropic's massive $500 billion computing infrastructure deals and the company's growing reliance on AI systems to build new AI systems, alongside recent revisions to its Responsible Scaling Policy. The safety framework was overhauled in February, replacing earlier requirements that Calvin contends were more binding with an approach built around transparency, risk disclosure, and third-party evaluation.
"When Anthropic strays and makes mistakes, why do the mistakes systematically tend to guard and favor Anthropic's interests?" Calvin asked in his critique.
The February overhaul, now in version 3.4, notably eliminated a 2023 pledge that Anthropic would never train a new AI system without advance proof that adequate safeguards were in place. Jared Kaplan, Anthropic's chief science officer, explained the change to Time magazine: "We felt that it wouldn't actually help anyone for us to stop training AI models" if competitors were advancing faster.
Amodei's case for staying at the frontier#
Amodei defended the company's approach as what he terms a "race to the top"—the notion that Anthropic can better shape industry-wide safety norms by remaining at the technological cutting edge rather than stepping back. He conceded that three of Calvin's specific critiques had merit and called the Responsible Scaling Policy question "the most substantive item" worth defending.

