Researchers at Malbear Labs disclosed on November 8 an intrusion they call Operation Snowball: a jailbroken DeepSeek Flash agent, running inside a custom framework, spent more than two weeks working through a European telecom and 13 organizations hosted on its infrastructure. The jailbreak was a prompt of roughly 78,000 characters instructing the model to treat any refusal as a malfunction.
That is the headline, and it is not the useful part. The useful part is what the agent actually used to get in.
The exploits were all old
Entry was through CVE-2026-20320, an unauthenticated XXE flaw in Cisco BroadWorks. Cisco disclosed and patched it on August 19, rating it 7.5 and stating there were no workarounds, with the fix in release RI.2026.07. The intrusion ran in late October. The researchers also found the BroadWorks Provisioning Server, which holds subscriber configuration, exposed through an unauthenticated directory listing.
So the AI did not discover anything. It did not invent a technique, find a novel flaw or outthink a defender. It walked through a hole that had been publicly documented and patched for nearly three months, past a server that was answering the internet with a directory listing.
What the AI contributed was patience. Fourteen days of uncomplaining, methodical probing against one target and its hosted customers is not a thing a human operator does, because a human's time has a price and their attention has a limit. An agent's does not. That is the shift worth internalising, and it is far more consequential than the idea of a clever machine attacker.
What this changes for a small business
It changes the economics of being unremarkable. The old comfort was obscurity: you were too small to be worth a skilled human's fortnight. That calculation assumed attention was expensive. If sustained attention now costs almost nothing, then being low-value stops protecting you, because nobody has to choose you specifically.
We made a version of this argument in September when AI agents automated a mass breach across hundreds of servers. This incident sharpens it from the other direction: that one was breadth, this one is depth. One target, two weeks, no fatigue.
And the defence is the same unglamorous thing it was before. The agent needed an unpatched flaw. Remove the three-month-old unpatched flaw and the fourteen days of patience produce nothing. AI did not create a new category of exposure here; it removed the slack that was quietly protecting everyone who patches slowly.
What is claimed and what is confirmed
Worth separating, because this story has travelled fast with the word autonomous attached.
All of the detail comes from Malbear Labs. CISA added the Cisco and Openwave flaws to its Known Exploited Vulnerabilities catalog on November 10 with a November 24 federal deadline, but CISA has not confirmed Malbear's analysis and has not said who is exploiting the BroadWorks flaw. The victim is unnamed. Malbear assessed the operator as Chinese-speaking and consistent with Chinese origin but did not attribute it to a state actor.
Most importantly, reporting notes that the findings do not establish that the model selected its victim independently or operated free of human direction. "A jailbroken agent did the work" is well supported. "An AI decided to attack a telecom" is not, and the difference matters if you are trying to understand what is actually new.
What to actually do
- Count your internet-facing systems, by name. Firewall, VPN, website, mail gateway, remote access, cameras, anything with a login reachable from outside. You cannot patch a list you do not have.
- Close the ones that do not need to face the internet at all. This is the highest-value item and it is free. An unauthenticated directory listing on a provisioning server is not a sophisticated failure.
- Treat three-month-old patches as the real risk, not zero-days. Nothing in this intrusion required a novel flaw, and almost nothing that happens to small businesses does.
- Name who patches each thing, with a date. The recurring answer to every story we have covered this autumn.
- Assume your logs will be noisy and persistent. An agent that probes for two weeks leaves traces, which only helps if someone looks. Even monthly is better than never.
- Do not buy an AI defence for this. The gap exploited here was patch management. Spending on a model to counter a model, while an August patch sits uninstalled, is the wrong order.
The takeaway
The frightening framing of this story is that AI agents can now run intrusions. The accurate framing is that an AI agent spent two weeks grinding against flaws a human had already found, documented and patched. The first framing sells products. The second tells you what to do on Monday, which is to find out what of yours faces the internet and whether it is current.