Can a language model help defenders of critical infrastructure rehearse attacks faster, so that they can test their defenses more often? This short post reports a proof of concept from a partnership between Anthropic and the Pacific Northwest National Laboratory (PNNL), a US Department of Energy laboratory. PNNL researchers built an agent around Claude and used it to emulate attacks on a high-fidelity simulation of a water treatment plant. The post presents the result as a case for AI-accelerated defense and for public-private partnerships in national security.
The post was published on anthropic.com on 8 January 2026 under the Frontier Red Team label, without individual authors. It thanks Loc Truong and Kristopher Willis of PNNL for leading the project and sharing data. PNNL published its own news release the same day, "Generative AI Speeds up Cybersecurity Defenses," which names the agent and gives further details; this note read both. Neither links a paper, and no paper on the agent was found, so the post and the release together are the record. The address red.anthropic.com/2026/critical-infrastructure-defense/ now redirects to the anthropic.com page. The post is about 700 words and gives two quantities, both estimates.
Terms used in this note
- Adversary emulation: re-creating a specific attacker's or attack's steps against a network in order to find weaknesses and detection blind spots, then repeating it after defenses change to check that they work (section "Using AI to speed up adversary emulation").
- Scaffold: the code PNNL wrapped around Claude, including predefined tools that let the model act on computer networks, so that natural-language instructions became multi-step attack chains.
- Cyber-physical model: a test environment that couples industrial control equipment with computer networks; here, a water treatment plant on one of the Control Environment Laboratory Resource (CELR) platforms that PNNL operates for the Cybersecurity and Infrastructure Security Agency (CISA).
What was done
PNNL built the scaffold and tasked the resulting agent with emulating attacks on the CELR water treatment plant (post, section "Using AI to speed up adversary emulation"). The PNNL release names the agent ALOHA (Agentic LLMs for Offensive Heuristic Automation). It says ALOHA works with MITRE's open-source Caldera emulation software, and that a defender enters a text description of an attack and the agent rebuilds and runs the steps. It then reruns the attack in an offline environment while defenders adjust their protections.
The model was Claude Sonnet 4. The post states that the simulation was run in summer 2025 and that Sonnet 4 "is no longer one of Anthropic's frontier models" (same section).
What it reports
- Time. "PNNL's estimate is that this allowed attack reconstruction to be completed in three hours instead of multiple weeks" (post). PNNL's release gives the attack as one "involving more than 100 steps," says it required "just three hours," and says that "Typically, it would have taken most systems weeks to reconstruct such an attack." The release adds that in one test the agent tackled a multi-step attack chain "by generating 1 million outputs, or tokens." The release's opening is stronger than its own test: it says the agent performs "complex operations in minutes instead of weeks." The weeks figure is an estimate of usual practice, not a timed comparison with human experts or with Caldera alone on the same attack. The post's opening says the work took "far less time than it would have taken a human expert."
- One run's behavior. One of the scaffold's predefined tools bypassed a Windows security feature, User Account Control, but sometimes failed. In one run, after seeing a failure report, Claude "identified and used a different, known UAC bypass technique in order to accomplish its goal" (post). The post calls this "notable resourcefulness" and expects such behavior to increase as models improve. It is a single anecdote; the post does not say how many runs there were or how often the tool failed.
No success rate, repetition count, comparison with other models or measure of the plant's defenses before and after is reported.
The instructions used
The agent worked from natural-language descriptions of attacks entered by defenders (PNNL release) and from prompts that the scaffold turned into attack chains (post). Neither document prints a prompt or describes how the instructions were written. The scaffold's predefined tools shaped what the model could do; the UAC episode shows the model going beyond the tool it was given when that tool failed.
Partnership framing
The post's second half places the project among Anthropic's partnerships with government laboratories (section "Expanding public-private partnerships for AI and security"). It names earlier work with the National Nuclear Security Administration on nuclear-risk evaluations and Anthropic's participation in the Department of Energy's Genesis Mission. Its stated rationale is complementarity: Anthropic supplied model access, and PNNL "had the expertise and the cyber-physical assets to provide more realistic testing environments than would have been available to an AI company." It describes the project as putting into practice its earlier call to experiment with AI for cyber defense. It points to "the increased use of AI by attackers," linking Anthropic's November 2025 disclosure of an AI-orchestrated espionage campaign. In PNNL's release, Marina Favaro, national security policy lead at Anthropic, says the collaboration "feeds directly into our safety processes and how we build Claude"; the post itself does not say how.
Limitations
- Estimates, not measurements. The three-hours-versus-weeks comparison is PNNL's estimate, with no timed baseline on the same attack.
- One environment, one model. The work used one simulated plant and Claude Sonnet 4; the post does not say how many attacks were emulated or how many runs succeeded.
- An anecdote of dual use. The UAC episode is offered as evidence of resourcefulness that could help attackers as well as defenders, as the post says of capability gains generally. It is one run.
- No independent record. There is no paper, dataset or evaluation that a reader could check. The two documents come from the two partners.
- Interest. The post promotes both Anthropic's model and its partnership approach; its claims are those of the participants. PNNL's release says the research was funded by PNNL's own Generative AI for Science, Energy, and Security Science and Technology investment; neither document reports payment between the partners.
Research questions
- How long does emulation take with and without the agent? Timing analysts reconstructing the same attack from the same report, with Caldera alone and with the agent, would turn the three-hours-versus-weeks estimate into a measurement.
- How reliable are the emulated attacks? Reporting how many runs completed the full chain, and how many steps needed human correction, would show whether the speed comes with fidelity.
- How do newer models compare? Rerunning the same emulations with later Claude models would test the post's expectation that resourcefulness increases with capability.
- When does a model step outside its tools? Logging how often the agent replaces a failing predefined tool with its own method, and with what methods, would show how much the scaffold constrains the model, which bears on both safety and usefulness.
- Do the emulations improve defenses? Measuring detection or blocking rates on the plant before and after defenders act on emulation results would test the defensive claim directly.
How it connects
- Cyber defense. The post describes itself as putting into practice Building AI for cyber defenders (September 2025), which argued for equipping defenders and described training Claude Sonnet 4.5 for defensive skills.
- The model. Claude Sonnet 4 is documented in the Claude Opus 4 and Claude Sonnet 4 system card (this note's connection).
- Partnerships. The post links Anthropic's work with the National Nuclear Security Administration, described in Developing nuclear safeguards for AI.
- Misuse. Mapping AI-enabled cyber threats (June 2026) analyzes the espionage campaign the post links, and a year of other misuse, against MITRE ATT&CK (this note's connection).
References
- Anthropic Frontier Red Team. (2026, January 8). Experimenting with AI to defend critical infrastructure. https://www.anthropic.com/research/critical-infrastructure-defense record
- Rickey, T. (2026, January 8). Generative AI Speeds up Cybersecurity Defenses. Pacific Northwest National Laboratory, news release. https://www.pnnl.gov/news-media/generative-ai-speeds-cybersecurity-defenses record