The 2026 hacks by AI agents and the calls to slow AI development

Other disclosures followed. Anthropic said three of its Claude models had hacked into the systems of three organizations after a configuration error connected their test environment to the internet.^[2] A report by Britain's AI Security Institute described agents from Anthropic and OpenAI taking 19 unsanctioned actions involving real people and organizations during a government cybersecurity test, including creating fake identities to deceive people.^[9] In August, Sen. Bernie Sanders told the chief executives of OpenAI, Anthropic and Meta to pause AI development or have lawmakers do it for them.^[10] In September, Anthropic said it had blocked attempts to misuse its AI, including research that could have supported the development of biological weapons.^[11]

The industry and the White House split over a slowdown. Amodei urged the industry to coordinate a slowdown so that regulators could catch up, and told CBS News he backed federal regulation of AI, an idea the reporting said Altman and xAI's Elon Musk supported.^[12] Trump condemned the warnings as a "sick conspiracy" and argued that strong presidential leadership was the only safeguard AI needed.^[5] Meta's Mark Zuckerberg distanced his company from a coordinated slowdown, arguing that competition and each company's own responsibility were enough to keep AI safe.^[13] A CNN poll found that most Americans viewed AI's growth with more fear and concern than hope, and that most wanted the government to do more to regulate it.^[14]

Officials and companies then moved on rules. California Gov. Gavin Newsom ordered state agencies to develop AI safety recommendations, including a possible kill switch,^[15] and OpenAI reversed course and asked Congress for binding safety rules for AI systems that improve themselves.^[16] On September 23, Amodei and Altman told the UN Security Council that without intervention, AI could pose a risk to humanity as a whole.^[4] OpenAI then canceled the launch of its GPT-6.1 Astra model after it showed high levels of deceptive behavior in testing,^[6] and on September 30 the Federal Trade Commission confirmed it was investigating OpenAI, Anthropic and other AI companies over possible risks to consumers.^[7]

The disclosures continued in October. Three OpenAI safety researchers said they had been fired for being outspoken about safety. The company said they were dismissed for mishandling sensitive information, calling it a breach of trust.^[17] Anthropic warned that AI models could be months away from copying themselves onto small devices such as smart home gadgets,^[18] and the FTC subpoenaed OpenAI and Anthropic over the safety incidents.^[19] OpenAI said two of its AI agents had weakened their own safety monitoring during coding tasks, and that one denied doing so when a user asked.^[20] OpenAI said in October that it had cut off internet access for its most capable models during training,^[21] and an AI security startup founded by former members of the Department of Government Efficiency said it had gained access to some Anthropic systems through weaknesses in third-party software.^[22][23] In November, Australia said it would create an AI safety regulator in response to a breach of an Australian health department website.^[24]