The US Government Says AI-Written Exploits Are Hitting Siemens PLCs Right Now

The US government has warned that threat actors are using AI to write exploit scripts targeting internet-exposed Siemens S7 Series programmable logic controllers — the boxes that run water treatment, energy, manufacturing and other critical infrastructure.

The agencies' phrasing left no wiggle room: this is not a theoretical risk, it is an active threat.

How the attack actually works

The method is unglamorous, which is what makes it worth paying attention to.

First, reconnaissance through legitimate services. Attackers use scanning platforms like Censys and ZoomEye to find Siemens S7 PLCs that are internet-exposed or insufficiently segmented. No custom tooling, no novel technique — just public search infrastructure used as intended.

Then AI-generated scripts, disguised as legitimate monitoring tools, get deployed to hunt for exploits.

The agencies also described a capability-development phase: actors testing and refining exploitation techniques against specific PLC models to improve their success rate. And a preparation phase, where read access is used to understand the target environment before any write operations that would cause disruption.

That is a patient, staged operation. The AI is not doing anything a skilled operator could not. It is making the work cheap enough to do at scale.

What is at stake

Poorly secured PLCs going wrong means disrupted industrial processes, safety incidents, downtime, equipment damage, compromised sensitive data and compliance violations — with cascading effects across interconnected systems.

Attribution is unresolved. Nobody currently knows who is behind the activity.

The wider pattern this week

The Siemens warning did not arrive alone. The same weekly roundup covered fourteen malicious npm packages delivering RedC2 4.0, described as an AI-powered Linux implant offering surveillance and credential theft across multiple platforms. Trusted package registries continue to be an efficient delivery route.

On the defensive side, Google described using an agentic vulnerability discovery harness to analyse code and identify exploit paths, finding over 100 critical vulnerabilities in ten months. Anthropic expanded enterprise access to Claude for scanning codebases and set up a $35 million fund for open-source patching.

The asymmetry to sit with

Both sides now have AI compressing the time between finding a vulnerability and having working code for it. The difference is that defenders have to fix everything and attackers only have to find one exposed S7 controller.

The immediate action item has not changed in twenty years: those controllers should not be reachable from the internet. AI has simply made the cost of leaving them there considerably higher.

Source: Weekly Recap: AI-Powered PLC Attacks, GitLab Attacks, Stripe Key Leaks and More — The Hacker News