How hackers use AI agents to breach 100+ company websites

One hacker's AI agents got into websites of 100+ companies. How attackers pair open-weight models with open-source AI pentest tools, and 3 things to do now.

One hacker, more than 100 companies. AI agents now do much of the work that used to take a skilled team, and open-source tools put them within reach of almost anyone. In this 2-minute video I explain how attackers put the pieces together, and what to check first.

Key takeaways

  • One operator, 100+ companies: Reuters reported a single hacker whose AI agents got into websites of more than 100 companies, for about $5,000 in tokens.
  • Speed: Microsoft reported a China-linked group that, in some cases, reached privileged cloud access less than three hours after first targeting a victim.
  • The tools are public: less-guarded open-weight models, abliterated versions with their refusals removed, and open-source AI pentest frameworks.
  • Fix the basics, then test continuously: know your exposure, put controls in front of it, and keep testing.

What happened

In October, two reports showed what AI-driven attacks look like in practice.

One hacker, more than 100 companies. Reuters reported that a Chinese-speaking operator pointed autonomous AI agents at more than 1,200 websites and got into sites belonging to more than 100 companies, including Deere, SharkNinja and Cummins. Snorre Fagerland, a threat researcher at NorForge, found the activity through exposed control panels for Cairn, an open-source AI problem-solving platform. He estimated the campaign used roughly $5,000 worth of tokens, about $50 per successful hack. SharkNinja said the attack hit a standalone product support site and that the threat was contained. (Reuters, via the Honolulu Star-Advertiser)

Privileged cloud access in under three hours. Microsoft Threat Intelligence described Storm-0940, a China-linked group that used an AI agent framework called ARTEX against more than 2,250 Microsoft 365 tenants. Its entry point was credentials stolen in a ConnectWise breach. In some cases, the group gained privileged cloud access less than three hours after first targeting a victim. Microsoft said the activity did not come from a compromise of its own systems. (CSO Online)

Why AI-driven attacks got cheap

The most capable cyber models, such as GPT-6 Cyber and Mythos, are gated. Getting access takes a long verification process.

Open-weight models are a different story. Models like GLM-5.3 and DeepSeek 4.1 are less guarded, and because their weights are public, anyone can download and modify them.

Abliterated models go a step further. Someone takes an open-weight model and removes its refusal behavior, so it almost never declines a task. Versions like these are published openly on model hubs.

Open-source AI pentest frameworks supply the rest. Projects such as Shannon, PentAGI and ARTEX give a model tools and a plan, so it can scan, probe and exploit on its own. ARTEX is the framework Microsoft says Storm-0940 used.

Put an unguarded model inside one of these frameworks and you get an automated attacker. It works around the clock, and in the Reuters case it cost about $50 per successful hack.

What to do now

  1. Know your exposure. List everything an unauthenticated outsider can reach: websites, APIs, login pages, remote access tools and cloud consoles. Attackers' agents scan all of it, so you should know it first.
  2. Put controls in front of it. Move internal tools behind a VPN or zero-trust access, put a web application firewall in front of public apps, require MFA, and alert on scanning from outside. Storm-0940 started with credentials stolen from a third-party tool, so review the credentials your vendors and remote-management tools hold.
  3. Test continuously and fix what you find. Attackers now run their tests around the clock. Run yours continuously too, fix what you find, and rerun the test to confirm the fix works.

Video transcript

This transcript is lightly edited for readability.

One hacker, more than 100 companies. If you open your feed recently, you probably saw the different AI-driven attacks that are hitting companies out there. Microsoft reported an attack where, in under 3 hours, hackers got privileged access to a victim's cloud environment. One hacker used AI agents to hack into more than 100 companies, including SharkNinja. So what's actually happening, and how is it happening?

We all heard about GPT-6 Cyber and Mythos, very capable models. But very gated. It's very hard to get access to them. You need to go through a very long verification process. But there are alternatives out there. GLM 5.3 and DeepSeek 4.1 are way less guarded. They will let you do way more, and potentially also help you try to run cyberattacks.

Now, it's not stopping there. Since those models are open weights, anyone can just download them and modify them. And there are some versions out there, the abliterated model versions, where someone took the same GLM 5.3 or DeepSeek 4.1 and reduced their guardrails. Those versions will almost never refuse any task, so attackers can use them to conduct attacks.

Now, add on top of that the open-source penetration testing frameworks out there, such as Shannon, PentAGI and ARTEX. You can integrate them with those models, and you get a powerful automated hacker.

So what should we do? We should probably start with our perimeter: our publicly facing resources, what external unauthorized users can access, and how we can protect them. Maybe move some of the resources behind a VPN, or make sure that we have a WAF in place. Make sure we have preventive controls and detection capabilities in place whenever something or someone is trying to scan our network from the outside world.

We also need a process of continuous security verification and remediation. We want to keep testing our own environments, find those issues, and mitigate them before attackers find them.

Questions

How did one hacker break into more than 100 company websites?

According to Reuters, a single Chinese-speaking operator pointed autonomous AI agents, run through the open-source Cairn platform, at more than 1,200 websites and got into sites belonging to more than 100 companies, including Deere, SharkNinja and Cummins. The researcher who found it estimated the campaign used about $5,000 worth of tokens.

What is an abliterated model?

An abliterated model is an open-weight language model that someone modified to remove its refusal behavior. It almost never declines a request, including requests the original model would refuse, which makes it useful to attackers.

What are Shannon, PentAGI and ARTEX?

They are open-source AI penetration testing frameworks. Connected to a capable model, they can plan and run parts of an attack on their own. Microsoft reported that the China-linked group Storm-0940 used ARTEX.

How can a company defend against AI-driven attacks?

Start with what an unauthenticated outsider can reach, put access controls and detection in front of it, and test your own environment continuously so you find and fix issues before an attacker's agents do.

Keep reading

  1. Prompt injection examples and how to stop indirect attacks

    How prompt injection and indirect prompt injection steer AI agents, four documented examples, and the layered controls that limit the damage.

    Daniel Begimher

    Testing
    9 min
  2. Five days, five security incidents that broke the internet.

    PHP2Shell is broadly exploited, LiteLLM 1.107.0 shipped malware (49,000 downloads), and OpenCode, Railway and eMudhra put keys at risk. What to do now.

    Daniel Begimher

    Notes
    5 min
  3. How to get an AI agent approved: 12 questions reviewers ask

    What security, risk and compliance teams ask before they approve an AI agent in a regulated company, and the evidence that answers each question.

    Daniel Begimher

    Standards
    9 min

Stop collecting findings. Start closing them.