100 Tech Giants Warned That AI Is Making Cyberattacks Easier. They Forgot to Mention Who Will Stop Them.
More than 100 companies signed an open letter about AI-powered hacking. A veteran security educator read it closely and found the plan had no people in it.
Key points
- More than 100 technology companies, including OpenAI, Microsoft, and Google, signed an open letter in late summer 2026 warning that AI is making sophisticated cyberattacks cheaper and more common.
- On August 19, 2026, five US federal agencies confirmed that attackers used AI-generated hacking scripts against industrial controllers at water treatment plants, power stations, and chemical facilities.
- Siemens confirmed no new flaw in its controllers was exploited: attackers simply walked through doors left open by old, unfixed configuration mistakes.
- Britain's RUSI defence think tank published research the same day arguing that criminals adopt new tools only when those tools make money, not simply because the tools exist.
- Every defensive action the open letter recommends requires a trained human to carry it out, yet the letter names no people, no funding figures and no deadlines.
Over 100 technology companies co-signed an open letter warning that artificial intelligence (AI, meaning software that can learn and make decisions on its own) is about to make sophisticated cyberattacks dramatically cheaper and far more common. The signatories included OpenAI, Anthropic, Google and Microsoft. The letter said defenders have "a limited window to strengthen cyber defenses."
James Lyne, chief executive of the SANS Institute (a major cybersecurity training organisation), read it twice. The second time, he went looking for who would actually do the work. He came up empty, and wrote about it in Dark Reading.
What is actually happening right now?
The threat is real. On August 19, 2026, five US federal agencies documented attackers using AI-generated exploitation scripts (code written by AI to probe and break into systems) disguised as ordinary monitoring software. The targets were Siemens S7 controllers, the industrial computers that run water treatment plants, power stations, and chemical facilities. We'd already tracked the broader exposure problem: in August we reported that Forescout counted 4,407 industrial controllers reachable from the open internet, including a cluster in towns recently hit by water-system attacks.
Building that kind of attack tool once required deep specialist knowledge. AI has removed that barrier.
Siemens issued its own response and made a telling admission: no new flaw in its controllers was found. Attackers used new techniques aimed at old, unfixed configuration mistakes. The door was already open.
| Key fact | Detail |
|---|---|
| Letter signatories | 100+ companies including OpenAI, Microsoft, Google, Hugging Face |
| Federal agency alert | August 19, 2026 |
| Systems targeted | Siemens S7 industrial controllers |
| Siemens finding | No new software flaw; old misconfiguration exploited |
| RUSI assessment published | Same day as the open letter |
Should ordinary people be worried?
Yes, in a measured way. The water plants and hospitals named in the letter serve real communities. But the most effective fix the letter recommends has nothing to do with AI: take the internet-facing controllers off the public internet. A well-staffed security team can do that on a Tuesday afternoon.
For a rural water utility run by one exhausted engineer juggling five jobs, it may as well be written in a foreign language. That gap is a staffing gap, not a technology gap.
What is missing from the plan?
Lyne's core complaint is simple: every recommendation in the letter is a verb (fix, deploy, share, scrutinise), but there's no named person, no budget figure and no deadline attached to any of them. A plan full of verbs with no subjects isn't yet a plan.
He offers three things he'd have liked to see. Invest in the engineers already working at facilities, because 15 years of site knowledge beats any fresh hire. Make "hands-on support" mean actual committed hours from named practitioners, not just access to a software tool. Judge AI-powered defence tools by whether a two-person hospital security team can use them under pressure, not by whether they make a large corporate team marginally sharper.
His sharpest aside: we're told in one breath that these AI models are too unpredictable to be held accountable when they cause harm, and in the next that they should be trusted to defend critical infrastructure. Both can't be true at the same time.
OpenAI did announce a $1 billion program called Daybreak for Frontline Defenders to subsidise AI security tools for utilities and public bodies, as we covered on September 3. Whether that commitment survives the next funding cycle without named practitioners and firm dates attached is exactly the question Lyne is asking.
If you work at a hospital or a small utility, push your leadership to ask not just what security tools they're buying, but who on staff is trained to run them when something goes wrong.
