Breaking

Anthropic CEO Says AI Industry 'Lied' on Risks, Presses Labs to Slow Capability Gains

September 13, 2026 at 17:59 EDT

Foundation Models
Policy & Regulation
AI Agents

Anthropic chief executive Dario Amodei said the artificial intelligence industry "lied to people about the fact that this technology had risks," arguing that an exponential rise in capability is a warning to slow frontier progress so safety work can catch up. In a CBS Sunday Morning interview with Jo Ling Kent, conducted at the company's San Francisco headquarters, Amodei said he had not "fully just appreciated what it would actually be like when the progress was as fast as it was."

September 13, 2026 · Anthropic

The industry long 'lied' on AI risks

Dario Amodei says exponential progress is a warning to slow capability gains so testing and alignment can catch up—not a reason to panic or shut the field down.

Time to act, drawn to scale

Amodei's swarm-risk window versus the extra alignment time he says could greatly cut the chance something goes seriously wrong. Columns are proportional (about 2×).

6–12 mo

Agent swarm could take over the internet

1–2 yr

Extra time sought for alignment and tests

Takeaway: A short buffer, used well, is the core of the pacing case.

6–12

Months until a misaligned swarm could seize the internet, Amodei warns

$100B+

Potential damage he cites from a persistent botnet takeover

1–2 yr

Extra time he says could greatly reduce serious-risk odds

5–10 yr

Horizon he still cites for curing most major diseases

Three-step pacing plan

STEP 1

Embedded evaluators

Employee-like access for outside testers. Anthropic committing now.

→

STEP 2

Democratic coordination

Shared standards and limits among labs in democracies.

→

STEP 3

Global coordination

Hardest step: agreements that include China, with verification.

Backing

Altman, Musk, and Hassabis endorsed pacing. OpenAI said it will match evaluator access. Some researchers called cooperation overdue.

Pushback

Critics call it a regulatory moat, doubt China will join, and say self-policing and third-party audits are not enough.

The fork he draws

Build it the right way, risk stays low. Build it the wrong way, risk goes high.

Testing

Every new model generation must be properly tested.

Kill switch

Could be a good idea; a complete ban is not, he says.

Oversight

Joint democratic governance, not one company or one state.

Continue reading

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