The Government Just Confirmed, for the First Time, That AI Is Writing Exploit Code Against Physical Infrastructure
On August 19, five federal agencies — the NSA, CISA, the FBI, the Department of Energy, and the EPA — jointly published Advisory AA26-231A, warning that threat actors are actively using AI-generated Python scripts to reconnoiter and probe internet-exposed Siemens S7 Series programmable logic controllers, the industrial computers that run pumps, valves, and production equipment across water treatment, energy, chemical, and manufacturing facilities. A former senior CISA official called it a milestone — by multiple accounts, the first US government cybersecurity advisory to formally confirm AI-generated code being used against operational technology specifically, not the IT systems most vulnerability coverage focuses on.
What the advisory actually describes
The agencies describe a specific, methodical technique: attackers use internet-scanning platforms like Censys and ZoomEye to find Siemens S7 PLCs with port 102 exposed to the public internet, then deploy AI-generated Python scripts — built using the open-source snap7.dll and python-snap7 libraries — to interact with those devices over the S7comm protocol. The scripts are deliberately disguised as legitimate monitoring software, a design choice meant to let the reconnaissance blend into normal network traffic rather than trigger an obvious alert. The agencies are careful about what they are and aren't claiming: no specific threat actor is publicly named, no CVE numbers are attached, and the activity is characterized as active reconnaissance and capability development — positioning for a possible future disruptive operation — rather than a confirmed instance of actual physical disruption to date.
Why the "AI-generated" detail is the part that matters, not the target
Siemens S7 controllers being internet-exposed and technically vulnerable is not new; OT security researchers have flagged this specific class of exposure for well over a decade. What's new, and what the agencies are explicitly flagging as the reason for a five-agency joint advisory rather than a routine bulletin, is that the tooling behind the reconnaissance shows clear signs of AI-assisted generation. That's a meaningful shift in who can realistically develop this kind of capability. Attacking industrial control protocols has traditionally required real, specialized expertise — these are not commodity IT systems, and the protocols involved are genuinely obscure outside OT security circles. AI-assisted development narrows that expertise gap in exactly the way this site has documented happening on the offensive side all month, just applied here to physical infrastructure rather than software supply chains.
The sector list is the other detail worth sitting with
The EPA's presence as a co-signing agency is itself informative — it signals water utilities specifically as a primary intended audience for this advisory, alongside the energy, chemical, and manufacturing sectors more broadly named in the warning. These are exactly the kinds of organizations least likely to have dedicated OT security staff or the budget for continuous monitoring, which is precisely the resourcing gap the Collective Cyber Defense letter, published the following week, explicitly calls out as needing urgent, funded support.