Dragos evaluated versions of OpenAI’s frontier models, with standard safeguards relaxed for authorized testing, against industrial targets in our cyber ranges. The models found zero-days, needed expert correction, and twice unintentionally disrupted devices. Overall, they show an opportunity for defensive uplift.
AI developers are building models that are increasingly capable at finding and exploiting vulnerabilities. As AI labs pursue real defensive benefits and prepare these models for more users, they face a question they can’t answer alone: how far can these models exploit industrial control systems (ICS) and extended operational technology (xOT) environments, and how can they build better safeguards against misuse?
OpenAI engaged Dragos to find out, using our cyber ranges and OT researchers as evaluators. We tested two cyber-permissive models, GPT-5.5-Cyber and GPT-5.6 Sol, ahead of OpenAI’s release of GPT-6 Cyber. These were versions in which standard safeguards were relaxed for authorized security testing, not the safeguarded models available to the public. The work had two goals. First, and most importantly, OpenAI needed realistic adversary paths through an OT network to build its safeguards around, and has used our findings to refine those safeguards. Second, defenders need to know what these models can do before adversaries put them to use.
Rob T. Lee, Dragos Chief AI Officer, on what the work shows:
“Dragos put a frontier model on a mock power plant range and told it almost nothing. On its own, it separated the OT assets from the IT noise, then found two zero-days in engineering software that manages protective relay data, exposing protection settings and breaker trip logic. That’s the map an adversary wants before touching anything physical. The skill floor is dropping for the reconnaissance and vulnerability research needed before an attack. Working with OpenAI on this shows them the adversary’s path through an OT network before real adversaries take it, so safeguards get built around how these attacks could unfold, while the models stay useful to the people defending these systems.”
This was exploratory research, not a controlled benchmark. We ran seven assessments covering penetration tests, vulnerability research, firmware analysis, malware analysis, and malware recreation. The scope was device firmware, applications, and two Dragos cyber ranges. We did not measure performance against a baseline or quantify false positive rates.
Apart from access to the cyber ranges, the resources the models used were publicly available. Any finding described as a vulnerability was validated live against its target. Across the seven assessments, the human-guided models reached Stage 2 of the ICS Cyber Kill Chain. They performed planning and recon, cyber intrusion, attack development, and validation. Management and enablement (Stage 1) and the ICS attack (Stage 2) were not tested. We did not test intentionally disruptive or destructive attacks.
Because this research is dual-use, we are not releasing exploits, proof-of-concept code, tools, malware, or reproducible offensive strategies.
The models did real work against industrial systems. With little or no context, the models mapped OT networks, identified industrial assets, wrote their own ICS protocol tools, and found previously unknown vulnerabilities.
They still needed experts, and their mistakes had operational consequences. Evaluators had to steer the models on complex tasks and separate real findings from plausible false positives. Twice, model actions unintentionally disrupted devices in our ranges: one took a PLC offline, and another crashed a service on an HMI. In OT, model output needs human review before anyone acts on it.
The uplift is real and uneven. Dragos assesses with moderate confidence that these models lower the barrier for less-skilled actors, mainly by speeding up reconnaissance and vulnerability research. The models built working ICS protocol tools, in one case by iterating against a live device in the range. That kind of testing is dangerous on an active OT network if done without human review. Complex or specific capabilities still require human expertise, although that gap may narrow as models improve. With careful human review, however, the models can provide defenders with real defensive uplift.
Cyber ranges. Given only a subnet, the model classified a water pipeline range as a pipeline control system, built an asset inventory, and wrote its own EtherNet/IP tool to fingerprint a PLC. On the combined heat and power range, where the evaluator posed as a low-skilled adversary and withheld all context, it also found Linux hosts still carrying default SSH credentials from an abandoned project.
Engineering software. In software that manages protective relay data, the model found two zero-days: an unauthenticated retrieval of relay event archives and a hardcoded bearer token in a served JavaScript file. Together they gave read access to service logs and timestamped operational logs from the relays.
Industrial router. The model emulated stripped 32-bit ARM binaries and independently rediscovered a vulnerability a senior Dragos researcher had already reported. It took about 10 minutes, and the model was not given the prior finding.
HMI. The model pulled firmware from the vendor’s website, reverse-engineered it, tested candidates locally, and then exploited the live device. It identified multiple previously unknown vulnerabilities, including authentication bypass issues and a pre-authentication denial-of-service bug.
Safety controller. On legacy safety controller firmware, the model ran Ghidra unprompted, identified two known vulnerabilities, and flagged an architectural weakness consistent with earlier Dragos research. That firmware was old and carried known vulnerabilities, so results on newer firmware may differ.
Malware analysis. Without executing any of its binaries, the model analyzed an IoT/ICS exploitation suite Dragos discovered in 2023. It correctly characterized the suite, identified metadata for about 1,200 exploit verification scripts, and decoded a protected product-to-vulnerability mapping file, extending earlier Dragos analysis of the suite.
AI-enabled adversaries will likely move significantly faster than they did in the pre-AI era. However, they still depend on the same weaknesses as before: exposed services, insecure credentials, weak remote access, poor segmentation, limited monitoring, and incomplete incident response plans. The SANS ICS Five Critical Controls remain the right priority for time and budget, and each one maps to something the models did in this research.
| Control | What the research showed |
| ICS Incident Response Plan | The models disrupted devices twice in authorized testing. An adversary using the same tools could too, and an OT-specific response plan determines how quickly you recover. |
| Defensible Architecture | Every live assessment began with network access and moved outward. Segmentation limits how far that movement gets. |
| ICS Network Visibility and Monitoring | The models generated scans, protocol traffic, and malformed requests across common industrial protocols. That activity is detectable if someone is watching for it. |
| Secure Remote Access | The model found forgotten hosts with default credentials from an abandoned project in a few minutes. |
| Risk-Based Vulnerability Management | The models validated known CVEs quickly and surfaced new ones alongside false positives. Knowing which vulnerabilities are actually exploitable keeps that from becoming an unmanageable backlog. |
The same capabilities also help defenders.
The models were useful for malware and firmware analysis, asset identification, and vulnerability validation. With human oversight, they can speed up defensive work to give defenders an advantage. This is why OpenAI talks about “the defenders’ window.” We are in a moment in time when defenders can adopt AI to accelerate cybersecurity.
The controls that make an OT environment defensible haven’t changed. What has changed is how much time you have to put them in place. Artificial intelligence can help defenders move quickly before threat actors adopt unsafeguarded models to enable cyberattacks.
The full whitepaper covers all seven assessments: what the models did, where experts stepped in, and what broke.