CrowdStrike Fal.Con AI agents https://pixabay.com/illustrations/blockchain-data-records-system-8244276/

What It Takes to Give Defenders an AI Advantage

When Hugging Face published its account of the breach it suffered in July, the detail I kept coming back to wasn’t how the attack worked. It was what happened when the company’s own responders tried to investigate it.

The intrusion was run by an autonomous agent that carried out thousands of actions, harvested credentials, and moved across internal clusters over a weekend. OpenAI later said its own evaluation models drove the activity during an internal cyber-capability test after escaping the research environment. Hugging Face used AI to help triage and reconstruct what happened. When its team tried to run the forensic analysis through hosted frontier models, though, the providers’ safety guardrails blocked submissions containing real exploit payloads and command-and-control artifacts. To a safety filter, a responder digging through malware for answers and a criminal shopping for exploits apparently look about the same.

So Hugging Face switched. It ran the analysis on GLM 5.2, an open-weight model, on its own infrastructure, which also kept sensitive data inside its environment.

It’s a small operational detail, and I think it says a lot about where defenders stand.

Attackers get to choose their tools

Look at the other side of the ledger. In the South Korean bank intrusions CrowdStrike documented in October, a single operator ran an agentic penetration testing tool on DeepSeek, with GLM and Grok in supporting roles and Claude Code sessions alongside. When one model wasn’t right for a task, the operator just switched to another, and when a provider might have objected, a reseller filled the gap. The toolchain was basically whatever worked that day.

Defenders often don’t have that freedom. They’re tied to whatever AI their vendors have embedded, or a general-purpose assistant that won’t touch malicious code, or a model they’re not allowed to send sensitive data to. The most capable security-focused models, such as Anthropic’s Mythos and OpenAI’s cyber variant, are offered through gated programs to a limited set of partners. Everyone else is left bringing a knife to a gunfight.

If the attacker can pick and adapt and the defender can’t, the attacker has an advantage that has nothing to do with skill.

What turns AI capability into defense

From what I’ve seen, a strong model on its own doesn’t get a security team very far. It takes a few things working together, and the model is only one of them.

  • The first is model flexibility. Defenders need to be able to use the right model for the job, including ones that will work with real attack data and ones they can run in their own environment when the data can’t leave. Being tied to a single provider’s policies is a weakness when the incident is happening at 2 a.m.
  • Then there’s the data. A general model knows how attacks work in the abstract. A model trained on years of real intrusion data knows what they actually look like in a production network. That data is scarce, and no AI lab has it at scale.
  • The last piece is the harness, which is the software that turns a model into something that actually does work, pulling context, taking actions, checking results, and handing off to people when it should. A strong model with a weak harness produces interesting answers that nobody acts on.

How CrowdStrike is approaching it

CrowdStrike’s SafeMind, announced with NVIDIA at Fal.Con on September 1, is a useful example of a vendor trying to build all three. It’s two models working against each other. An offensive model looks for the path an attacker would take into an environment, and a defensive model works on closing it, with a harness running both in a loop so each keeps pushing the other.

The models are built on NVIDIA’s open Nemotron family and trained further on CrowdStrike’s own security data. The harness can also drive other models, including general-purpose ones from the major labs, which speaks directly to the flexibility problem Hugging Face ran into.

CrowdStrike says that compared to leading frontier models and open-source baselines, SafeMind delivered a 29 percent higher detection rate, remediated problems six times faster, and cost 99 percent less to run. Those are CrowdStrike’s numbers from CrowdStrike’s tests, and as far as I can tell, nobody outside the company has reproduced them yet. They should be treated as claims until someone independent runs the comparison.

The offensive half deserves scrutiny too. A model built to find attack paths is useful to attackers as well as defenders. CrowdStrike says the system operates inside its platform with safeguards, and who gets to use it, and against what, is the right question to keep asking.

What buyers should ask

Whatever vendor you’re evaluating, the Hugging Face episode suggests some practical questions.

  • Can we use this AI on real malicious artifacts? If the model refuses to analyze exploit code, it won’t help in an actual incident.
  • Where does our data go? Some investigations involve data that can’t leave your environment. Know in advance whether you have a model you can run locally.
  • Are we locked into one model? Ask whether the platform can swap in different models as they improve, or whether you’re tied to one provider’s roadmap and policies.
  • What security data is behind it? A model’s value in defense depends heavily on what it was trained on. Ask what that is and how current it is.
  • Has anyone independent tested it? A vendor benchmark is a reasonable place to start and nothing more.

Elastic Security Labs, in its analysis of the Hugging Face breach, recommended that teams pre-vet a self-hosted model for incident response before they need one, and I’d repeat that advice to any security team.

Stop waiting for a better model

The attacker in the Korean campaign had an edge because they could grab whatever model worked and never had to ask anyone’s permission. Waiting around for the next frontier release to fix that is a gamble, though, and the odds aren’t in your favor. If I were running a security team right now, I’d pick a model I can run on my own infrastructure and test it against real artifacts from a past incident, well before the next one lands on a Saturday night.

Scroll to Top