Anthropic Urges Slower AI Progress Amid Risks
Anthropic is publicly arguing for the ability to slow frontier AI development even as it restricts access to its most capable internal models. In a blog post by Anthropic Institute lead Marina Favaro and co-founder Jack Clark, the company warned that "full recursive self-improvement," in which AI advances itself without human input, is approaching, and that the industry "has a gas pedal, but it doesn't have a brake pedal." Clark told the BBC and CNN that any credible slowdown would require multiple frontier labs across countries to pause under verifiable, shared conditions, comparing the challenge to Cold War arms-control stabilization. Clark also said roughly 80% of Anthropic's own coding is now done by its Claude model. Separately, Anthropic has restricted its Mythos model, which it described as strikingly capable at computer-security tasks, to a small consortium of major technology companies rather than releasing it broadly.
The argument
In a blog post by Marina Favaro, who leads the Anthropic Institute, and co-founder Jack Clark, Anthropic warned that "full recursive self-improvement," where AI systems advance themselves without human intervention, is approaching, and that the field lacks the means to stop if needed. Clark's framing, repeated to the BBC and CNN, is that the industry "has a gas pedal, but it doesn't have a brake pedal."
What a slowdown would require
Anthropic is explicit that a unilateral pause is not the proposal. Clark argues any credible slowdown would need multiple well-resourced labs at or near the frontier, across multiple countries, agreeing to stop under the same conditions, with each able to verify the others have actually done so. He likened the coordination problem to Cold War efforts to stabilize the nuclear arms race.
Capability context
The appeal lands alongside evidence of fast-rising automation. Clark said roughly 80% of Anthropic's coding work is already performed by Claude. The company has also restricted its Mythos model, which it described as strikingly capable at computer-security tasks, to a curated group of major technology firms rather than releasing it, citing misuse risk.
Why it matters for practitioners
For teams responsible for risk, governance, and deployment, the message is that gated, partner-only model access and heavier red-teaming are becoming standard frontier-lab practice. Absent enforceable international coordination, industry self-governance and national regulation will shape the near-term rules, raising compliance and oversight burdens for deployers.
Key Points
- 1Anthropic is calling for a verifiable, multi-lab 'brake pedal' on frontier development as recursive self-improvement nears, signaling more governance scrutiny ahead.
- 2Clark's claim that ~80% of Anthropic's coding is done by Claude underscores rising model autonomy and the need for stronger code review, CI/CD, and auditing.
- 3Restricting the high-capability Mythos model to a vetted consortium shows frontier labs gating dangerous capabilities, foreshadowing more partner-only access programs.
Scoring Rationale
A frontier lab publicly arguing for a verifiable, coordinated slowdown, while simultaneously gating a highly capable security model, is a notable governance signal for practitioners responsible for risk and deployment. It is significant and well-sourced but reflects a position and proposal rather than an enacted policy or technical breakthrough, keeping it below the industry-shaking tier.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems