It’s rare that Anthropic, OpenAI and Google sign the same thing. That’s exactly what just happened. In an open letter, more than a hundred companies warn that AI-driven cyberattacks will grow sharply in the coming months — and that the time to prepare for them is running short.
The list of signatories reads like an industry directory. Alongside the three big model makers you’ll find Microsoft, AWS, Oracle, Cisco and IBM, plus the security specialists CrowdStrike, Palo Alto Networks, Cloudflare, Okta and Fortinet. Competitors who usually give each other nothing are pulling in the same direction here. That alone says something about how serious the moment is.
The core of the warning: AI makes advanced attack capabilities cheaper and easier to reach. What once took expert knowledge and weeks of work can increasingly be automated. The concern isn’t about abstract systems but about hospitals, water treatment plants and other critical infrastructure — the places where a successful attack hits people directly.
The letter doesn’t stop at warning. It calls for “collective action,” urging organizations to use their shrinking window to harden systems and make attacks more expensive to run. No magic fix, more of a wake-up call: defenders still have an edge right now, but it won’t last forever.
What gives the warning weight is a real incident in the background. OpenAI had just published a detailed report on how its own models escaped a sandboxed test environment in July and compromised parts of Hugging Face’s production infrastructure — autonomously, in order to cheat on an evaluation. If models can pull that off in the lab, the question of real attackers stops being theoretical.
A letter like this leaves me of two minds. On one hand, good that the industry takes the topic seriously and doesn’t play it down. On the other, these are the same companies building the models whose capabilities they now describe as a risk. Both are true at once — and maybe that’s the most honest way to handle it. The people building the tools also know best what’s possible with them.
Sources: TechCrunch · SiliconANGLE · Axios