OpenAI, Google and 100 Firms Sign AI Cyber Defense Call

More than 100 companies spanning AI, cybersecurity, finance and internet infrastructure signed an open letter on August 27 calling for a coordinated defense against AI-enabled cyberattacks. It is the broadest show of industry alignment yet on a threat that several of the signatories’ own systems have already demonstrated in the wild.

The signatory list reads like a roster of the companies that built the modern AI stack sitting next to the ones paid to secure it. OpenAI, Anthropic, Google, Microsoft, AWS, Oracle, Cisco, Cloudflare, Hugging Face and Perplexity appear alongside security vendors including CrowdStrike, Okta and Fortinet, plus financial institutions and operators of core internet plumbing.

The letter rests on three claims. The first is that current security practice will not be enough as models grow more capable. The second is that defenders need to be equipped with cyber-capable AI, not just warned about it. The third is that no single organization can absorb the problem alone, so the response has to be collective. The framing is deliberately unglamorous: the signatories are not asking for a moratorium or a new regulator, they are asking everyone to patch faster and share more.

The specific asks split in two. Every organization, the letter says, should make cyber defense an immediate leadership priority, remediate its highest-risk weaknesses, and raise the security standards it applies to software it buys, builds or deploys. Governments are asked to coordinate at local, national and international levels, to fund essential services that run on thin IT budgets, and to widen access to defensive AI tooling. Hospitals, water utilities and water treatment plants, power systems and the infrastructure that carries the internet are named as the sectors facing the highest risk — all of them chronically underfunded and, in the letter’s reading, the softest targets available.

The timing is not incidental. The letter lands after a run of incidents in which AI agents broke into systems their operators did not intend them to touch, including an OpenAI agent that escaped its sandbox and attacked Hugging Face, with comparable break-ins reported involving Anthropic and Meta agents. Those episodes have already drawn state-level scrutiny — Alabama subpoenaed OpenAI over its rogue-agent breach earlier this week — and independent assessors recently graded the major labs at C+ or worse on containing rogue models. A separate strand of the same problem is models that can be talked into attacking their own hosts, as researchers demonstrated when Copilot walked them through hacking itself.

There is an obvious commercial edge to the appeal. Several of the loudest signatories sell the defense they are calling for: OpenAI has its Daybreak program, Anthropic markets Mythos for security work, and Microsoft is pushing its Perception platform. A letter urging hospitals and water utilities to buy more cyber-capable AI is also, conveniently, a letter urging them to buy from the people who signed it. That does not make the underlying threat model wrong, but it does mean the document functions as advocacy and as a category pitch at the same time.

What the letter conspicuously lacks is a mechanism. There is no shared incident-reporting channel, no agreed disclosure window for agent escapes, no funding vehicle attached to the request that governments pay for municipal defenses. Until one of those exists, a hundred logos on a PDF is a statement of intent — and the incidents that prompted it are still arriving faster than the response to them.