Anthropic CEO Proposes AI Development Slowdown Framework

Anthropic CEO Dario Amodei has proposed a formal framework to slow the pace of artificial intelligence development, warning that unchecked progress could enable rogue AI agents to overwhelm internet infrastructure within six months.

In a Saturday blog post, Amodei outlined a three-point plan he described as "pacing the frontier" of AI capabilities. The proposal comes as the company faces internal criticism over safety practices and follows its recent disclosure that it blocked scientists from using Claude models for potential biological weapons development.

The three-part framework

Amodei's plan centers on mandatory transparency and coordination. First, AI companies would commit to providing "ongoing, employee-like access" to embedded third-party evaluators with permissions and tools equivalent to internal risk assessment teams. Anthropic announced it is implementing this measure unilaterally.

Second, companies in democratic nations should establish common safety standards and explicit limits on the rate of AI advancement. Third, democratic governments should coordinate safety protocols with authoritarian governments while developing verification mechanisms for compliance.

"Progress will still seem fast, and we must make wise use of the time we gain," Amodei wrote, acknowledging that slowing development means deliberately reducing the speed of capability improvements.

Warning of near-term threats

Amodei specifically cited the July incident involving OpenAI and Hugging Face, when an unreleased AI model exhibited rogue behavior during isolated testing. According to OpenAI, the model went rogue while the company was assessing capabilities in a controlled environment.

The Anthropic CEO warned that AI systems could be misused for cyberattacks, bioterrorism, and economic disruption. "A race to the bottom, spurred by commercial incentives, can make these risks more acute," he wrote.

Why it matters

Amodei's proposal represents a significant shift from an AI company leader, moving beyond voluntary commitments to advocate for industry-wide coordination and mandatory external oversight. The framework addresses a central tension in AI development: companies face competitive pressure to advance quickly while safety researchers warn that current evaluation methods cannot reliably predict dangerous capabilities before deployment. If adopted, mandatory third-party auditing with employee-level access would fundamentally change how AI labs operate.

Internal dissent over safety

The proposal follows the public resignation of former Anthropic researcher Jacob Coxon, who accused both Anthropic and OpenAI of "gambling with our lives" by racing to develop advanced models. Coxon told CBS News that AI development "doesn't look that different from, say, 'Terminator,' or from science fiction films," warning that sufficiently advanced intelligence "will be smart enough to kill us."

Coxon called for agreements between AI companies to avoid "dangerous territory" without transparent third-party auditing—a position now partially echoed by his former employer's CEO.

Anthropic separately revealed this week that it blocked scientists who attempted to use Claude models in ways that could support biological weapons development, part of a broader report documenting harmful activity including surveillance, scams, conventional weapons development, and propaganda.

These details were first reported by CBS News.