Breaking

Anthropic CEO Dario Amodei Asks Frontier Labs to Slow Capability Gains Until Safety Catches Up

September 12, 2026 at 21:55 EDT

Foundation Models
Policy & Regulation
Security

Anthropic CEO Dario Amodei has called on frontier AI companies to slow how quickly they raise model capabilities, arguing that alignment, interpretability, and independent evaluation must catch up before more autonomous systems can cause catastrophic harm. He is not asking labs to freeze training. He wants a slower, verifiable pace—and Anthropic is opening its own systems to outside reviewers now.

September 12, 2026 · Anthropic

Pace the frontier — don’t freeze it

Dario Amodei asked AI labs to slow capability gains until alignment, interpretability, and outside reviewers catch up. Training would continue. The race, he argues, should be for safer systems, not just stronger ones.

6–12

months until a more capable misaligned swarm could, he warns, scale into an internet-wide botnet

1–2

extra years of focused safety work that he says could greatly cut the chance of a serious failure

steps: on-site evaluators, democratic standards, then harder global deals

3–5

years he says chip and distillation controls could widen the U.S. lead

Time windows, drawn to scale

Each block is 6 months. Columns use the lower swarm bound, the 12-month upper bound, and the 2-year upper bound on extra alignment time.

6 mo

Swarm risk, lower bound

12 mo

Swarm risk, upper bound

24 mo

Extra safety time, upper bound

Takeaway: The extra alignment time he wants is about twice the upper end of the near-term swarm-risk window he cites.

Three-step pacing plan

Step 1

Embedded evaluators

Employee-like access inside labs. Anthropic is doing this now.

→

Step 2

Democratic standards

Shared limits among labs, with a narrow antitrust waiver.

→

Step 3

Global coordination

From bio-weapon bans to a possible pause. Full pause called unlikely soon.

Support

Sam Altman said OpenAI will also grant independent evaluators employee-like access. Elon Musk backed the call. Hugging Face asked to join the reviewer program.

Pushback

Critics call extinction warnings a distraction and joint lab rules a form of capture. The rapid sequence of insider alarms and this plan is also being read as a coordinated push for tighter rules.

What outside reviewers would get

Publish the findings — even the unflattering ones

Anthropic says reviewers can report risk levels, incidents, practices, and which access they received or were denied. The company may redact security, legal, commercial, or third-party secrets — not a result just because it looks bad.

Desks, badges, and company laptops on site
Permissions largely comparable to internal risk teams
Potential damage cited if a swarm scales: hundreds of billions of dollars

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.