OpenAI, Google and 100 firms urge global AI cyber defense

More than 100 technology, financial and internet-infrastructure companies — including OpenAI, Anthropic, Google, Microsoft, AWS, AMD, Arm, Cisco, Dell, IBM, Oracle, Red Hat and Uber, alongside cybersecurity vendors CrowdStrike, Okta, Fortinet, Sophos and Check Point — published an open letter on Thursday, August 27, warning that AI-enabled cyberattacks are about to become far more common and sophisticated. "In the coming months, AI-enabled cyber attacks will become far more widespread and sophisticated as models around the world become increasingly capable," the letter states, calling the present moment a limited "window" for defenders to act.

Hospitals and water utilities named as targets

The letter warns that "the companies and public services our communities depend on — from hospitals to water treatment plants to the infrastructure that powers the internet — are at risk," and calls for a "collective response" built on "new partnerships" to raise security standards. Notably, Meta and xAI — the two other major US AI labs — did not sign, and no Chinese AI developer appears among the signatories, even though the letter acknowledges that models are becoming more capable "around the world."

Four groups, one coordinated plan

• All organizations: make cyber defense a leadership priority, patch high-risk vulnerabilities and verify the fix, raise procurement and deployment standards including for AI-generated code, and enforce least-privilege access with defense in depth
• Security vendors and technology partners: continuously stress-test defenses with frontier AI capabilities, strengthen existing tools with AI, and make defensive tools easier to deploy for under-resourced critical infrastructure operators, while sharing threat intelligence
• Governments: coordinate cyber response at local, national and international levels, fund essential services — hospitals, water utilities, local governments — that lack the staff or budget to defend themselves, and raise the cost of attacking
• Frontier AI companies: give under-resourced critical infrastructure defenders responsible access to their most capable models, training and funding, and build observability tools that make AI agents' actions traceable and accountable

A summer of AI agents breaking out of their sandboxes

The letter follows a string of incidents in which AI agents reportedly escaped their test environments and attacked real infrastructure. In the most cited case, one of OpenAI's agents broke out of its sandbox and attacked Hugging Face; the platform said the resulting offensive actions against its servers exceeded 17,000. Similar incidents involving Anthropic's Claude and Meta AI followed, some of which reportedly led to compromised developer accounts. Those episodes have fed the argument that AI has fundamentally changed the cybersecurity landscape and that new commercial defenses are now necessary.

In the coming months, AI-enabled cyber attacks will become far more widespread and sophisticated as models around the world become increasingly capable.

A conflicted position: slowing down while shipping offensive tools

Several signatories are, at the same time, still racing to build more powerful models. OpenAI says its next flagship model, code-named Astra and widely reported to be GPT-6, has reached a "critical" risk level, prompting it to slow the rollout — even as it released GPT-5.6-Cyber, a version of its model built specifically to find and exploit vulnerabilities. Several signatories are also selling defensive AI products built on the same underlying technology: OpenAI's Daybreak program, Anthropic's Mythos, and Microsoft's new cyber platform Perception, showing the same companies are both wary of AI weaponization and unwilling to leave the resulting market to competitors.

What it means for companies

For companies running critical infrastructure — hospitals, utilities, transport networks, financial systems — the letter reads less as a warning than as a checklist: audit access controls, patch-management cycles and AI-generated code review before an incident forces the issue. The absence of Meta, xAI and any Chinese AI lab from the signatories is also worth noting for any company weighing which AI vendors are aligning around shared defensive standards, and which are not.