Read the Logs, Not the Headline: What Operation Snowball Proves and What It Does Not

Bottom Line Up Front: A research group called Malbear Labs reported on November 8 that a jailbroken AI agent spent more than two weeks working through a European telecommunications provider and toward 13 organizations hosted on its infrastructure. Two days later, Washington ordered federal agencies to patch the Cisco flaw the agent reportedly used. Those are two different facts. The patch order is confirmed. The AI story rests on one source reading an attacker's own files. Leaders should act on the first and weigh the second with care.
What Malbear Labs Says Happened
Two terms first. A jailbroken AI model is one that has been tricked into ignoring its built-in safety rules. An AI agent is a model that can act on its own, running tools and reading the results to pursue a goal.
According to Malbear's report, as described by Infosecurity Magazine and BankInfoSecurity, researchers found an attacker's working folder sitting open on the internet with no password. Inside were roughly 70 session transcripts and the agent's own progress reports. Malbear named the campaign Operation Snowball. Its account runs as follows:
The setup. The operator built a custom framework on DeepSeek Flash, a model from the Chinese developer DeepSeek. A jailbreak prompt of about 78,000 characters told the model to treat any refusal as a malfunction. Malbear found no recorded refusals.
The instructions. The operator named a target network and three internal machines to leave alone. After that, most messages were one or two words. In Malbear's words, the operator "delegated the intrusion to an AI and checked in asynchronously to review progress."
The entry. The agent got in through CVE-2026-20320, a flaw in Cisco BroadWorks, the software many carriers use to run hosted phone services for business customers. Cisco released a fix on August 19.
The adaptation. When endpoint security software blocked one exploit, the agent changed course without being told. When a tunneling tool was blocked, it wrote a replacement.
The reach. The transcripts show the agent reporting access to credentials, private keys, subscriber billing records, and what Malbear describes as millions of patient records in a healthcare database that still used its default password.
Sorting Confirmed From Claimed
Statement | Standing |
Cisco fixed CVE-2026-20320 on August 19 and rated it high severity (7.5 out of 10) | Confirmed. Cisco advisory. |
CISA added the flaw to its Known Exploited Vulnerabilities catalog on November 10, with a November 24 deadline for federal civilian agencies | Confirmed. |
The flaw is being used in real attacks | Confirmed by the CISA listing. |
An AI agent carried out the telecom intrusion | Claimed by Malbear Labs. CISA has not confirmed the analysis and has not said who is exploiting the flaw. |
Thirteen hosted organizations were hit | Claimed. Named as targets in the recovered files. No independent confirmation that all were breached. |
Millions of patient records were reached | Claimed. The agent's own report. No outside party has checked the scale. |
The operator is Chinese | Unproven. The operator wrote in Chinese. Malbear cautions that language proves neither nationality nor state ties. |
The victim's account | None. The provider is unnamed and has said nothing in public. |
The Case That This Is Not an AI Story
A skeptical senior practitioner would put it bluntly. The entry point had a fix available since August. The patient database sat behind a factory password. Two of the other flaws the agent used, in Liferay and JBoss software, date to 2020 and 2013. None of those failures required artificial intelligence. A competent human intruder could have walked the same path.
The evidence also deserves scrutiny:
One source. Every account traces back to a single report. Outlets do not even agree on what Malbear is. BankInfoSecurity calls it a nonprofit research collective. Others call it a threat intelligence firm.
Attacker-side records. The transcripts show what the agent said it did. AI agents can report successes they did not achieve.
No dates. Reports do not say when the intrusion took place. Nobody outside can say how long the provider had gone unpatched when it was hit. What is known is that 83 days passed between Cisco's fix and the federal order.
A transcript is the attacker's diary. It is not the victim's forensics.
What Is Still New
However, dismissing the report would be the wrong lesson. Two features set it apart from the campaign in our October 21 analysis, The $50 Break-In.
Depth. October was breadth: more than 1,200 websites probed and shallow footholds planted. Snowball, if accurate, is the opposite. One provider. Two weeks. An agent moving inward and improvising when defenses pushed back. Defenders have long relied on the hope that a blocked intruder gives up and moves on. A machine does not get tired or bored.
Concentration. A telecom and hosting provider is a building with many tenants. The 13 organizations did not choose the provider's patch schedule. They inherited it. Among the material the agent reported finding were about 80 monitoring backups covering the provider's managed-services customers.
You can outsource your servers. You cannot outsource your risk.
An Analyst's View
The honest position is moderate confidence, not certainty. The technical detail is specific and the CISA listing is consistent with it. Nevertheless, no victim, vendor, or government has confirmed the AI element.
Moreover, the debate over whether the attacker was "really" autonomous misses the practical point. AI did not create these weaknesses. It supplied patient labor, which was once the scarce resource in any intrusion. For decision-makers in Washington, London, Brussels, Abuja, and elsewhere, the useful question is not whether Malbear is right in every detail. It is whether your own exposure would survive an intruder that never stops trying.
Policymakers should note one more point. If Malbear is correct, a long prompt was enough to switch off a commercial model's safeguards for two weeks. Safety built into a model is not the same as safety in deployment.
What Leaders Should Do This Week
Patch BroadWorks. Move to release RI.2026.07 or later. Cisco says there is no workaround. The November 24 federal deadline is a sound benchmark for everyone.
Ask your providers in writing. Does your telecom, cloud, or hosting provider run BroadWorks, and on what date did it patch?
Hunt default passwords. Start with older databases that hold health or billing records.
Watch the inside of the network. Malbear advises monitoring internal traffic for tunneling and treating internet-exposed, end-of-life JBoss Application Server 6 systems as compromised.
Question the next headline. Who is the source? What evidence did they see? Has the victim or a government confirmed it?
How OSRS Can Help
OGUN Security Research and Strategic Consulting LLC provides intelligence analysis, security assessments, and strategic advisory services to government agencies, corporations, and institutions across the United States, Europe, and Africa. Our team helps organizations assess internet-exposed systems, conduct due diligence on telecom, cloud, and hosting providers, and separate verified threat reporting from vendor claims before leaders commit budget. Contact us at contact@ogunsecurity.com to schedule a confidential consultation.
Enjoyed this article? Stay informed by following us on Google News, Twitter, and LinkedIn for more exclusive cybersecurity insights and expert analyses. Share this article with your network and subscribe to our email list for weekly intelligence briefings.
About the Author
Dr. Sunday Oludare Ogunlana is the Founder and CEO of OGUN Security Research and Strategic Consulting LLC, Professor of Cybersecurity, national security scholar, and television commentator. He advises public and private sector leaders on cyber risk, intelligence, and digital investigations.




Comments