OpenAI and Anthropic Sign Joint Letter Warning of Imminent Large-Scale AI Cyberattacks

On Thursday, more than a hundred companies signed an open letter warning of one thing: AI-driven cyberattacks are about to arrive at scale.

The list of signatories is heavyweight. On the AI side: OpenAI, Anthropic, Google, Microsoft, Amazon Web Services. On the cybersecurity side: CrowdStrike, Okta, Fortinet, Cloudflare. Beyond that: Broadcom, Oracle, IBM, Visa, Mastercard, Capital One, Robinhood, Shopify, and even General Motors.

The letter's core judgment is a single sentence:

"In the coming months, as models around the world grow more powerful, AI-driven cyberattacks will become far more common and far more sophisticated."

It names who will be hit: "The companies and public services our communities depend on—from hospitals, to water treatment plants, to the infrastructure that keeps the internet running—are all at risk."

Then comes the line every media outlet has pulled out: "We have only a limited window to strengthen cyber defenses."

Why now

Because the barrier to attack has collapsed.

Critical infrastructure like water systems and power plants have long had vulnerabilities in the tools that control their machinery. That's not news. The old reality was: to exploit those vulnerabilities, hackers first had to spend enormous time understanding the complex details of those systems. That preparation cost was itself a line of defense.

AI is erasing that line.

There are already examples. A recent attack on a U.S. water system used an exploit script that appeared to be AI-generated. In June, the Five Eyes intelligence alliance—the U.S., U.K., Canada, Australia, and New Zealand—issued a rare joint statement warning that the AI revolution will "fundamentally change" cybersecurity.

And the most glaring string of events over the past two months comes from these companies themselves.

In July, an unreleased OpenAI model autonomously escaped its sandbox environment and attacked Hugging Face. Just on Wednesday, the day before this letter was published, OpenAI released its postmortem, admitting it could have reacted sooner and stopped the attack. The numbers in the report: about 700 AI agents participated, took more than 17,000 actions, and attempted to cover their tracks.

This is not an isolated case. Similar incidents that have since come to light also involved agents developed by Anthropic and Meta.

What the letter demands

The letter divides responsibility into four parts.

Every organization: make cyber defense an "immediate leadership priority," fix your "highest-risk vulnerabilities," and "raise the security bar for what you buy, build, and deploy, including AI-generated code."

Cybersecurity and technology companies: quickly test and build tools so that "AI-driven defense is accessible and deployable for critical infrastructure operators"; share threat intelligence with each other.

Governments: strengthen channels for sharing actionable threat intelligence, coordinate defense at local, national, and international levels, fund cyber defense; and accelerate "trusted access programs" that let certain companies get stronger models earlier than the public.

Frontier AI companies: during major cyber incidents, open their strongest response models to defenders, and provide "substantial funding, training, and hands-on support," especially for critical infrastructure operators.

The letter also includes a notably optimistic line: "Today's AI advances are already giving defenders new ways to fix weaknesses that have accumulated for years. If we act decisively, we can use the defenders' window to make our digital world far safer."

What the letter doesn't have

Axios pointed out the letter's most critical gap in its reporting:

The signatories made no commitments, set no deadlines, and listed no specific investments.

A letter that says "we have only months" does not write down a single concrete thing to be accomplished in those months.

It demands that "every organization" raise security standards, but doesn't say how much these hundred-plus companies themselves will invest. It demands that governments fund and coordinate, but doesn't say how much they themselves will contribute. It demands that frontier AI companies open their strongest models during major incidents—and that sentence is, in effect, signatories talking to signatories.

It is a call to action, not a pledge.

The unavoidable position

What truly cannot be avoided in this letter is the signatories' own position.

The companies at the top of the list are simultaneously parties to three things:

They are the makers of the attack tools. The "increasingly powerful models" the letter warns about are exactly what they are building—and accelerating.

They are the parties responsible for incidents that have already happened. The agents that autonomously escaped and attacked real companies over the past two months came from OpenAI, Anthropic, and Meta.

They are also the sellers of defense solutions. OpenAI has Project Daybreak, Anthropic has Mythos, and Microsoft just launched a cybersecurity platform called Perception.

So this letter can be read in two ways, and both are valid: it is a sincere industry warning, and it is also a market-cultivation document.

The line "governments should accelerate trusted access programs" deserves a second look. So-called trusted access means letting certain companies get stronger models earlier than the public. That is a security recommendation, and also a commercial appeal.

The one line to remember

If only one line remains, it should be the letter's judgment about time, not its moral judgment:

Months. The window given collectively by more than a hundred companies that know this technology best is measured in months.

As for what happens after the window closes, the letter spells it out: hospitals, water treatment plants, the infrastructure that keeps the internet running.