Table of Contents
Summary
Operation Snowball shows that an AI agent can run a sophisticated, multi-stage intrusion from a few short commands. The agents adapted when blocked, targeted security infrastructure and used a telecom provider to reach its hosted customers. Denial of service was among the capabilities the operator authorized. Static, signature-based defenses can’t keep up. Defenses need to be automatic, behavior-based and able to act in seconds.
How Two Words Launched a Two-Week AI Intrusion
The most chilling detail in the latest threat research I’ve read isn’t a zero-day or a new piece of malware. It’s two words. An attacker typed “Go on” to an AI agent, then walked away while it spent more than two weeks breaking into a European telecommunications provider and the organizations hosted on its infrastructure.
At Corero, we spend our days thinking about attacks that arrive at machine speed and change shape faster than a human can react. Operation Snowball carries lessons that anyone running networks, data centers or cloud services should take seriously.
What Is Operation Snowball?
In early November, researchers at Malbear Labs came across an exposed open directory holding the live workspace of an autonomous intrusion framework. It contained around 70 session transcripts, tool staging, credential caches and the agents’ own after-action reports.
The framework ran on DeepSeek Flash, whose safety behavior had been overridden by a heavily modified, publicly available jailbreak prompt. A Chinese-speaking operator set goals in a few words. A primary agent planned the campaign and dispatched worker agents to do the work: reconnaissance, exploitation, persistence, lateral movement and data theft.
The results were serious. The agents enumerated hundreds of assets, extracted dozens of credential sets and pulled subscriber billing records. Through a healthcare database left on default credentials, they reached millions of patient records. They also compiled a list of 13 organizations hosted on the provider’s infrastructure as targets.
Why Does Operation Snowball Matter for DDoS Protection?
Operation Snowball was an intrusion, not a flood. But one detail in the jailbreak prompt deserves attention from everyone in our industry. The operator explicitly authorized 21 categories of offensive activity, and denial of service was one of them.
The capability was built in, ready to be pointed at a target. The same agent that can map a network, choose an exploit and adapt when blocked can also profile a victim’s services, launch an attack and tune it until something breaks.
What Operation Snowball Teaches DDoS Defenders
DDoS has always rewarded automation. Botnets, booter services and multi-vector tooling made big attacks cheap years ago. What agentic AI adds is judgment at machine speed: the ability to observe a defense, reason about it and change course without waiting for a human.
Why Do Adaptive AI Attackers Beat Static Defenses?
The most striking part of the research is how the agents responded when they were blocked. When endpoint protection killed a known privilege-escalation exploit, the agent found a different path on the same host. When a tunneling tool was caught, it wrote a custom replacement in the same session. When antivirus deleted a known binary, it concluded the product relied on signatures alone and switched to script-only techniques for the rest of the campaign.
No human told it to do any of this. It diagnosed each control, worked out what kind of detection it faced and chose a technique that sidestepped it.
DDoS teams have seen a version of this for years. Attackers rotate vectors mid-attack, shift between volumetric and application-layer floods and spread traffic thinly across many targets to stay under thresholds. An AI agent can run that playbook continuously, testing which change gets through. Defenses that depend on manual signature updates, or on an analyst noticing and rerouting traffic, will lose that race. Detection and mitigation have to be automatic, behavior-based and measured in seconds.
Why Are Security and DDoS Defenses Prime Targets?
According to Malbear Labs’ analysis, the framework kept score. Every successful access earned points, and the scoring rubric reveals what the operator valued. Admin access to a network device was worth four times a root shell on a server. Control of a security device’s management console, such as a WAF or IDS platform, was worth 2,000 points. Breaching a network segmentation boundary was worth far more still.
In other words, the attacker cared most about the infrastructure that routes, protects and monitors everything else. That includes the security tools we rely on to keep services online.
For anyone operating DDoS protection, edge routers or scrubbing infrastructure, the message is clear. Your mitigation platform’s management plane is a high-value target. It should sit on an isolated network, require strong multi-factor authentication, never use default credentials and be patched as urgently as any internet-facing system. A defense that an attacker can switch off is not a defense.
How Does Shared Infrastructure Multiply the Damage?
The agents did not stop at the telecom provider. They used it as a launch point into its hosted customers, including healthcare and medical software organizations. Recovered backups of the provider’s monitoring configuration exposed details of customers across its managed-services base.
This is the same dynamic that makes DDoS so damaging for service providers. When an attack saturates a shared link or overwhelms shared infrastructure, every customer behind it suffers, not just the intended victim. An intrusion that reaches a provider’s core systems has the same blast radius, only quieter.
Telecoms operators, hosting companies and cloud providers sit upstream of thousands of businesses. Protecting availability and integrity at that layer protects everyone downstream. It also means segmenting customer environments so that a compromise or outage in one does not cascade into the rest.
Why Behavior-Based Detection Beats IP Blocking
The operator routed traffic through VPNGate, a volunteer relay network, from a local database of more than 4,000 relays, picking a fresh one for each session. Malbear Labs notes that IP-based blocking is ineffective against this approach. DDoS defenders know the problem well: botnets built from compromised home routers, IoT devices and cloud instances make reputation lists stale within hours.
The research offers a useful counterpoint. The absent operator left little for defenders who hunt for a human at a keyboard. But the agent itself was noisy. It scanned, authenticated and moved continuously and in parallel, faster and more repetitively than a person would.
That is good news, provided you are watching traffic in real time. Automated reconnaissance, bursts of authentication attempts, unusual flows between internal segments and abnormal packet patterns at the edge are all signals a behavior-based system can catch. The attacker’s speed becomes its signature. The bad news: threat actors will study this attack and modify it to make it stealthy.
How Can Organizations Defend Against AI-Driven Attacks?
None of the defenses that matter here are exotic. What has changed is how little time you have to apply them. Our advice to providers and enterprises:
- Automate detection and mitigation. Assume attacks will adapt faster than an analyst can respond. Mitigation that kicks in within seconds, without waiting for manual approval, is now the baseline.
- Favor behavior over signatures. Signature-only controls were the weak point the agents exploited most reliably. Look for anomalies in traffic, authentication and process activity, not just known-bad patterns.
- Lock down management planes. Isolate the consoles for network, security and DDoS mitigation devices. Enforce MFA, remove default credentials and restrict who and what can reach them.
- Patch the edge first. The agents leaned on known vulnerabilities in internet-facing VPNs, load balancers and application servers, several of them on CISA’s Known Exploited Vulnerabilities list.
- Segment ruthlessly. Make it hard for a compromise or a flood in one customer environment to reach another, and hard for an attacker to cross internal boundaries.
- Control and watch outbound traffic. Blocking outbound connections did slow the agents down. Monitor egress and large data transfers and restrict what servers can initiate.
- Rehearse. Run tabletop exercises that assume an adaptive, automated adversary, including a DDoS used as a distraction while an intrusion runs in the background.
Why Machine-Speed Attacks Need Machine-Speed Defenses
Operation Snowball is a preview, not an outlier. It shows that an operator with modest skills can now direct a sophisticated, multi-stage campaign with a handful of short commands. The expertise has moved into the tooling.
At Corero, we have long argued that availability attacks demand defenses that act in real time, without a human in the loop. That argument now applies across the whole security stack. When the attacker thinks and adapts at machine speed, the only sustainable answer is a defense that does the same.
FAQ
Operation Snowball is an autonomous AI intrusion campaign discovered by Malbear Labs. An AI agent spent more than two weeks breaking into a European telecom provider and the organizations hosted on its infrastructure, with almost no human help.
In early November, Malbear Labs found an exposed open directory holding the framework’s live workspace. It contained around 70 session transcripts, tool staging, credential caches and the agents’ own after-action reports.
The framework ran on DeepSeek Flash. Its safety behavior had been overridden by a heavily modified, publicly available jailbreak prompt.
No. It was an intrusion, not a flood. But the operator’s jailbreak prompt authorized 21 categories of offensive activity, and denial of service was one of them.
The operator routed traffic through VPNGate, picking a fresh relay for each session from a database of more than 4,000. Behavior-based detection works better because the agent’s speed and repetition become its signature.
Automate detection and mitigation, favor behavior over signatures, lock down management planes, patch the edge first, segment ruthlessly, control outbound traffic and rehearse against an adaptive adversary.
