Published: February 25, 2026

What happened: Anthropic has replaced its two-year-old Responsible Scaling Policy with a more flexible framework it calls a "Frontier Safety Roadmap." The key change: the requirement to pause training more powerful AI models if their capabilities outstripped the company's safety controls has been removed. The new framework describes its guidelines as "public goals" rather than binding commitments.

Why it matters: The pausing commitment was Anthropic's flagship safety promise — the hard constraint that set it apart from less cautious rivals. Its removal signals that competitive pressure has pushed the company to prioritise market position over self-imposed limits, even as it continues to position itself as the AI industry's conscience.

Wider context: The announcement lands the same week that Defense Secretary Pete Hegseth gave Anthropic CEO Dario Amodei a Friday deadline: roll back AI safeguards or lose a $200 million Pentagon contract and face effective government blacklisting. Anthropic says the policy shift is unrelated to that ultimatum, but the timing is difficult to ignore. The company maintains two non-negotiable limits — AI-controlled weapons and mass domestic surveillance.

Background: Anthropic was founded by former OpenAI researchers specifically over AI safety concerns and has described itself as the AI company with a "soul." Its original policy was designed to create an industry "race to the top" — hoping rivals would adopt similar safeguards. The company now acknowledges that hasn't happened, and that unilateral restraint while competitors advance no longer makes strategic sense.

Anthropic ditches its core safety promise in the middle of an AI red line fight with the Pentagon — CNN Business

Singularity Soup Take: Watching a company founded on "we're the responsible ones" quietly swap hard commitments for aspirational goals — while simultaneously battling Pentagon pressure — raises a pointed question: if not Anthropic, then who?

Key Takeaways:

• Pause rule removed: The core requirement to halt training more powerful AI models if capabilities outstripped safety controls has been eliminated — this was the defining feature of Anthropic's previous policy.
• Non-binding goals: The new Frontier Safety Roadmap frames its guidelines as "public goals we will openly grade progress towards" rather than hard commitments Anthropic is bound to honour.
• Pentagon deadline: Defense Secretary Hegseth gave Anthropic a Friday deadline to drop AI safeguards or lose a $200 million government contract and face a de facto blacklist — Anthropic says the policy shift is unrelated.
• Two red lines remain: Anthropic still refuses to enable AI-controlled weapons or mass domestic surveillance, citing the technology's unreliability and the absence of legal frameworks governing such use.
• Industry race failed: Anthropic acknowledges its original policy was designed to inspire similar commitments from rivals — a "race to the top" that never materialised, making its own restraint counterproductive.