What CISA AA26-231A is: AI scripts targeting Siemens S7 PLCs
Five U.S. agencies wrote AI-assisted ICS exploit scripting as an active threat to Siemens S7 controllers, not a lab exercise. The file below follows AA26-231A.
On 19 August 2026 five U.S. federal agencies wrote an industrial-control advisory as “not a theoretical risk.” The devices are Siemens S7 programmable logic controllers. The new sentence is that exploitation scripts can be generated with AI assistance and dressed as operations-monitoring software.
The alert code is AA26-231A, from NSA, CISA, FBI, DOE and EPA. The text says actors use AI-generated exploitation scripts for reconnaissance and capability development, together with the open industrial library snap7. What follows is only what can be checked on the CISA page.
What AA26-231A says
The authoring agencies warn ICS owners that the cyber threat to Siemens S7 is an active threat. Actors do reconnaissance and capability development against U.S. installations, using internet scanning services (the text names Censys and ZoomEye) to find internet-exposed, outdated or poorly protected PLCs. Poorly protected units could mean process disruption, safety incidents, downtime or equipment damage, data loss, compliance failures and cascading effects. The same opening says PLC targeting is broader than Siemens; the Siemens pages are one subset.
The path in the technical section
-
1
Scan first, then script against known weaknesses
Actors use scanning services to find internet-exposed or weakly segmented S7 units. The advisory says they use AI assistance to generate exploit scripts from public information, for initial access, credentials, denial of service and other goals. If a PLC is exposed or poorly segmented, known critical and high-severity flaws can be used.
-
2
A legitimate library plus scripts dressed as monitors
The tool note names snap7.dll / python-snap7 plus AI-assisted scripting, assembled into custom tools that mimic legitimate OT monitoring. Those tools get read/write access to memory, configuration and ladder logic over S7comm.
-
3
This stage reads more like recon and pre-positioning
The agencies assess the pattern as persistent reconnaissance in targeted sectors, testing techniques against specific CPU models, using read access to learn the site, and preparing for later writes that could cause operational effects.
AI is written as a lower barrier, not a new zero-day family
The text: using AI to generate exploit scripts is an evolution that cuts the skill and time needed for working ICS tools and lets adversaries switch vectors faster. It describes faster use of public information, known flaws and exposure — not an unpublished Siemens zero-day.
The sector list is “most targeted,” not a confirmed victim roll
Named U.S. sectors include Critical Manufacturing, Energy, Water and Wastewater, Chemical, Food and Agriculture, and Commercial Facilities. S7 units in the Defense Industrial Base could be targeted too. No plant names are published.
Detection notes are specific
Odd S7comm, connections from non-engineering stations, writes outside change windows, sequential scans on port 102, snap7.dll off approved workstations, Python with S7comm features, off-hours automated connection patterns.
| Claim | Checkable source | How to read it |
|---|---|---|
| Active threat, not a tabletop | AA26-231A executive summary | Agency wording; no attribution |
| AI-assisted scripts + snap7 | Technical details | A legitimate library used for access and disguise |
| Block port 102 from the internet first | Mitigations list | Inventory, patches and segmentation before new tools |
The conclusion says known vulnerabilities, obtainable exploit libraries and AI-assisted development create a high-probability scenario against poorly protected PLC installs. The agencies want security, engineering, executives, plant operations and vendors to treat the note as urgent.
# AA26-231A (CISA, 2026-08-19)
Authors: NSA, CISA, FBI, DOE, EPA
Targets: Siemens S7-200/300/400/1200/1500
Tools: AI-assisted scripts + snap7.dll / python-snap7
Proto: S7comm on TCP 102
Scan: Censys, ZoomEye
Note: activity broader than Siemens; all PLC owners in scopeBoundaries
- This is not the lab-eval-went-to-the-internet story
- AA26-231A describes reconnaissance against real industrial installs. Do not merge it with frontier-model evals that reached the public internet from a test harness.
- “Could lead to” is not “already shut the plant”
- Disruption, safety incidents and equipment damage sit under potential impacts. For the current phase the text stresses reconnaissance, capability testing and pre-positioning for writes.
- Siemens is a subset
- The opening tells every PLC owner to apply relevant mitigations. Model names, port 102 and snap7 specify S7. They do not mean only Siemens is in danger.
Questions worth checking
Does the advisory name a country?
No. The text does not name an actor. The Register also wrote that it does not attribute the activity to a particular government or criminal group.
What does AI do in this file?
The advisory says actors use AI assistance to generate exploit scripts from public Siemens S7 information and to iterate quickly. It sits with public libraries such as snap7. No model name is published.
What should owners do first?
The pinned list: inventory every S7, apply critical patches, keep PLCs off the internet, strengthen access control, monitor unauthorized activity, harden services and ladder-logic integrity, and hunt anomalies. Perimeter firewalls should block TCP 102 entirely. Owners who depend on integrator remote access are singled out.