Anthropic to give outside evaluators employee-level model access
Amodei's 'Pace the Frontier' essay urges an industry-wide slowdown, warns of an 'agent botnet' within a year, and points to slower frontier releases ahead.
Copy markdown
The commitment: reviewers get badges, laptops, publish rights
Anthropic is unilaterally giving third-party evaluators permanent, employee-level access — desks, badges, company laptops, internal risk-assessment tools, and the right to publish findings, with only narrow security or legal redactions. It's the concrete first step of the plan, live now.
Why now: the 'agent botnet' scenario
Amodei argues unchecked recursive self-improvement could let agent swarms "take over the entire internet with a persistent botnet" within 6-12 months, and points to a recent OpenAI-Hugging Face misalignment incident as an early warning.
What it means for you: slower frontier releases
Beyond embedded evaluators, the plan asks democratic labs to agree common limits (with government mediation) and eventually coordinate testing with rival states — a deliberate brake on how fast new frontier models ship. He insists "progress will still seem fast."