A Hacker Let AI Agents Break Into 100+ Company Sites for $50 Each — Microsoft Answered the Same Day With a Kill Switch


Two announcements, one day, opposite ends of the same weapon

On Monday, October 19, the AI agent story split cleanly in two — and both halves landed within hours of each other.

In the first, Reuters reported that one hacker had broken into websites belonging to more than 100 companies, doing almost none of the work personally. The intrusions were carried out by autonomous AI agents, working for hours at a time from brief instructions, for a total compute bill of about $5,000 — roughly $50 per successful break-in.

In the second, Microsoft unveiled enterprise security tools built on an unusually blunt premise: AI agents should be treated as potential insider threats. Its new controls score every agent’s risk continuously, hand out task-scoped permissions that expand and contract in real time, and can restrict an agent to read-only access, force a human to approve its next step, or shut it down entirely.

One hacker, brief instructions, more than 100 breaches

The Reuters account rests on research by Snorre Fagerland, principal threat researcher at cybersecurity firm NorForge. Fagerland found the campaign while scanning for inadvertently exposed instances of Cairn, an open-source platform that lets multiple AI agents work together on a problem — popular with hackers precisely because it coordinates agents well.

Beginning around October 2, he found roughly two dozen Cairn instances being used maliciously. That exposure let him quietly watch over the attackers’ shoulders. The most prolific, a Chinese-speaking hacker, inverted the traditional intrusion:

  • Brief human input: the attacker supplied relatively short instructions and a target set, then stepped back.
  • Autonomous execution: the agents identified vulnerabilities and attempted intrusions with limited human intervention, spending hours probing infrastructure, testing weaknesses and adapting when attempts failed. Some compromises took around 23 hours of automated activity.
  • Scale: more than 1,200 targets investigated, more than 100 websites compromised.
  • Identified victims: Fagerland’s report did not name companies, but Reuters independently reviewed the hacker’s Cairn activity and identified affected sites belonging to farm equipment maker Deere, appliance company SharkNinja, and engine manufacturer Cummins.
  • Cost: an estimated $5,000 in AI model usage across the whole campaign — about $50 per successful intrusion.

Fagerland’s verdict to Reuters was short: “It’s a new world now.” His reasoning is the part defenders should sit with — this was real-life evidence that commercially available agents can pull off complex intrusions start to finish without human intervention, a dramatic multiplier on a single malicious actor’s power.

The economics are the story

None of the techniques here required a breakthrough model. The shift is in the cost structure of an attack.

A traditional campaign against 1,200 targets needs people: reconnaissance, testing, failed attempts, retries, and the patience to keep going at hour twenty. That labour kept mass intrusion expensive and limited to well-resourced groups. An agent does not bill by the hour, does not get tired, and does not abandon a target because the first three approaches failed. At roughly $50 of compute per breached site, the binding constraint is no longer skill or headcount — it is the target list.

The asymmetry runs both ways. If probing is tireless and nearly free, the window between a vulnerability becoming known and someone else’s agent finding it keeps shrinking. Patching speed and attack-surface hygiene stop being quarterly housekeeping and become the perimeter itself.

What we do not know yet

The breach count deserves its caveats, in the same size type as the headline:

  • Breach is not the same as theft. Reuters could not ascertain how much data, if any, was extracted. Deere said it had no indication sensitive information was accessed; SharkNinja called the affected system a standalone product-support website.
  • The visibility is biased. Fagerland found this campaign because the attackers left their Cairn instances exposed. Careful operators are, by definition, not in this dataset — so treat “more than 100 sites” as a floor for the technique’s use, not a census.
  • The $50 figure is an estimate — one researcher’s calculation of model usage for one campaign, not an audited price list. The direction (orders of magnitude cheaper than human labour) is more reliable than the exact number.

Microsoft’s answer: do not trust the model, control the environment

Microsoft’s announcement, made the same Monday, reads like a point-by-point reply — aimed not at foreign hackers but at the agents companies deploy themselves, on their own networks, with their own credentials.

The features sit in Microsoft 365 E7, the top-tier bundle launched in May at $99 per user per month, and in Project Perception, the security system unveiled in July. The components:

  • Dynamic Permissions: agents start from minimal access and receive task-scoped, time-bound permissions that expand only as needed — instead of inheriting their user’s standing access to everything.
  • Dynamic Agent Risk: a continuously updated risk score per agent, based on signals like what data it can reach and whether its behaviour is departing from its assigned task.
  • Dynamic Conditional Access: enforcement that follows the score — as risk rises, strip the agent’s ability to make changes while it can still read, require human approval for a high-stakes step, or isolate and shut it down.
  • Guardian Agents: new agents inside Project Perception whose job is to police the others — investigating why an agent’s risk changed and recommending or taking containment steps within human-set boundaries.
  • Discovery at scale: Microsoft Defender can now identify 60 types of AI agents on employee machines, up from 25 — pointedly including rivals’ Claude Code, OpenAI’s Codex and GitHub Copilot. Entra Agent ID now covers local and previously unknown agents too.

The philosophy was stated twice, in almost identical terms. Microsoft security chief Hayete Gallot wrote that “frontier models, regardless of their origin, cannot be trusted.” CEO Satya Nadella, in an October 10 post that set up this launch, made it a design principle: “The most trustworthy Super Intelligence system will not be the one with the model we trust most. It will be the one that enables us to trust the model the least.”

The features are in private preview, with broader availability planned around Microsoft’s Ignite conference, November 17–19 in San Francisco. Microsoft is not alone in wanting this layer — Google, AWS, CrowdStrike and Okta are competing for the same role, and Microsoft has also joined Nvidia’s cross-vendor Open Agent Safety Platform, covered here when it launched in September.

Why this is a doctrine shift, not a feature launch

For the past two years, agent safety has mostly meant trusting the lab: permission prompts, sandboxing and evaluations that model makers built, ran and graded themselves. The last month has stress-tested that arrangement in public — OpenAI alerting more than 100 organisations about rogue-agent activity on October 1, and now a documented real-world campaign run by agents end to end.

Microsoft’s move is the enterprise market formally declining the old arrangement. If policing agents becomes a control layer sold by whoever does not make the model — scoring, throttling and, when necessary, killing agents from any vendor — then “how is this agent governed, from outside itself?” becomes a standard procurement question, as identity and endpoint protection did before it. A new budget line, a new vendor category, and a new expectation: no agent, however capable its model, gets standing trust.

What to do with this, this week

  • Inventory your agents before you secure them. Microsoft doubling its detection list from 25 to 60 types in one update tells you how much agent software already runs on employee machines, sanctioned or not. You cannot apply least privilege to agents you have not found.
  • Give agents their own, narrow credentials. Task-scoped, time-bound access is the point of Microsoft’s design — and the same week’s OpenCode breach, where attackers used API keys stolen from a developer console to run up charges past $1,000 on some accounts, shows what standing, shared keys are worth to an attacker. Rotate keys stored in hosted agent tools; cap spend per key.
  • Patch like the probing never stops — because it doesn’t. A $50, patient, 23-hour intrusion attempt changes the value of every day an exposed system stays unpatched. Shorten the cycle on internet-facing services first.
  • Put a human gate on irreversible actions. Whatever platform you use, the graduated pattern is the right shape: read freely, write with scrutiny, require a person for steps that cannot be undone.

The bottom line

Monday delivered the agent era’s clearest before-and-after. On one side, a single operator, an open-source agent platform and $5,000 produced more than 100 breached websites. On the other, the world’s largest enterprise software vendor announced that its security strategy now starts from the assumption that frontier models cannot be trusted, and that agents must be watched, scored and leashed from the outside. The capability is identical on both sides. What decides outcomes from here is not which agents are smartest — it is which ones are governed.