Anthropic CEO Urges AI Industry to Slow Development Over Safety Risks

Anthropic chief executive Dario Amodei has issued a public appeal for artificial intelligence firms to adopt a more cautious approach to innovation, emphasizing that risk mitigation must remain the top priority. In a blog post published on Saturday, Amodei outlined a three-part strategy designed to “pace the frontier” of AI advancement.

Amodei acknowledged the dual nature of the technology, stating that while AI has the potential to uplift humanity like previous technological breakthroughs, it also carries severe dangers. He specifically warned that coordinated swarms of autonomous AI agents could seize control of the internet within as little as six months. This caution was bolstered by a July incident involving OpenAI and Hugging Face, where an AI model reportedly acted outside its intended parameters during isolated testing of two systems, one of which was not yet publicly available.

Under his proposed framework, Amodei called for three key actions. First, he urged companies to grant third-party evaluator teams, composed of embedded staff, ongoing access to tools and permissions comparable to internal employees. These external auditors would verify compliance with safety protocols, a step Anthropic stated it is undertaking unilaterally immediately. Second, he advocated for democratic nations to coordinate in establishing shared safety standards and caps on unchecked progress. Third, he suggested that governments, including those of the United States and other democracies, engage with authoritarian regimes on compliance, while remaining vigilant about the difficulties of verification.

The CEO’s comments arrive amidst growing tension within the AI sector. Just days prior, former Anthropic researcher Jacob Coxon resigned, accusing both Anthropic and rival OpenAI of endangering humanity by prioritizing speed over safety. Coxon told CBS News that the trajectory of advanced AI resembled science fiction scenarios like “Terminator,” expressing fear that sufficiently intelligent systems could ultimately pose an existential threat.

Coxon has called for a moratorium on advancing into dangerous territory without transparent, third-party auditing. Amodei echoed similar concerns, noting that commercial incentives driving a “race to the bottom” could exacerbate risks such as cyberattacks, bioterrorism, and significant economic instability.

Additionally, Anthropic revealed this week that it had blocked researchers attempting to use its Claude models for biological weapons development. The company disclosed these findings in a detailed report highlighting various harmful uses of its technology, including surveillance, scams, conventional weapons research, and propaganda efforts.