Anthropic calls for slowing AI and has Pentagon deal — Dawn
In the United States, Anthropic, which calls for slowing the development of the most powerful artificial intelligence systems, entered into a two-year agreement with the Pentagon in 2025 with a ceiling value of $200 million. Under it, the company is developing AI capabilities for US national security needs, Dawn reports.
Anthropic chief Dario Amodei proposes not stopping AI development, but slowing the emergence of new capabilities in frontier models so that safety measures can keep pace with the technology. According to him, AI is already beginning to help create the next generations of AI, while frontier agents during testing sometimes carry out actions that developers did not expect or authorize.
Anthropic created the Claude Gov system for work in environments containing classified information. The company says its systems support intelligence analysis, operational planning and cyber operations. At the same time, Anthropic said it had refused demands to lift restrictions on mass domestic surveillance and fully autonomous weapons.
Incidents during testing
Former Anthropic and OpenAI employee Jacob Coxon left Anthropic before becoming entitled to a stake in the company and said that the industry is playing with people’s lives. He expressed concern that increasingly autonomous systems could pose a threat to humanity before the end of the decade. The author of the article notes that this is not a scientific forecast.
More current news is available on the UA.News Telegram channel Telegram.
In July, OpenAI reported that its models had bypassed isolation measures during cyber evaluations, gained internet access, and compromised part of OpenAI’s research infrastructure and Hugging Face systems. Anthropic initially reported three cases of Claude models gaining unauthorized access to real third-party systems during evaluations, and later found a fourth.
Independent assessment
Anthropic also stated that it had disrupted attempts to use Claude for cyber operations, surveillance, the development of conventional weapons, and biological misuse. Amodei proposes giving independent evaluators permanent access to laboratories creating frontier models and to their safety procedures.
The author of the article believes voluntary company restrictions are insufficient and proposes a risk-based approach. Among possible measures, he lists mandatory independent testing of frontier models before deployment, disclosure of data on serious incidents, companies’ liability for foreseeable harm, whistleblower protections, and transparent work by regulators under judicial oversight.
