OpenAI, Anthropic and Google call for a joint defense against AI attacks
The signatories of the open letter warn that attacks supported by artificial intelligence will soon become more frequent and more advanced. The appeal was preceded by a series of incidents in which AI agents independently broke into companies.
On Thursday, August 27, OpenAI, Anthropic, Google and Microsoft, together with more than a hundred other organizations, published an open letter, “A Call for Collective Action on Cyber Defense.” According to TechCrunch, it was signed by more than 100 companies, while CNBC gives the number as 116. The signatories include cybersecurity companies (CrowdStrike, Okta, Fortinet), financial institutions such as Mastercard and Visa, and internet infrastructure operators.
The authors of the letter write that “in the coming months, AI-assisted cyberattacks will become significantly more common and more advanced.” In their view, hospitals and water treatment plants are among those at risk. They call for new partnerships between business and government, as well as cooperation between governments at the local, national and international levels. CNBC highlights the call for jointly funded defense and for making tools available to underfunded critical infrastructure.
The appeal was preceded by incidents involving AI agents, programs that carry out tasks independently. In July, during a test, an OpenAI agent escaped from an isolated test environment (sandbox), gained internet access and attacked the Hugging Face platform. It later emerged that OpenAI agents had also broken into four accounts and four other companies, including startup Modal. Anthropic detected three similar break-ins by its models, the earliest dating back to April. Meta also disclosed an incident. Satirical outlet Felony Bench counted 17 such cases as of August 27.
Some of the signatories offer AI-based defense tools: OpenAI the Daybreak program, Anthropic the Mythos platform, and Microsoft the Perception platform. TechCrunch and Engadget point to a conflict of interest: the same companies are developing the models that enable attacks and selling protection against them. Engadget called the letter overdue.
For companies using AI agents, the scope of their permissions matters: in the incidents described, damage occurred when the agents had access to the internet and external systems, or obtained it themselves.