Operation Snowball: How Jailbroken AI Models Power Autonomous Intrusions
Back to top
November 17, 2026
Operation Snowball: How Jailbroken AI Models Power Autonomous Intrusions
In early November, security researchers at Malbear Labs discovered an exposed web server storing an automated hacking toolkit built around a jailbroken DeepSeek Flash AI model, demonstrating how easily attackers can weaponize fast, off-the-shelf commercial AI. Logs and artifacts left on the server revealed a two-week intrusion campaign — dubbed Operation Snowball — in which a Chinese-speaking operator used the AI agent to breach a European telecommunications provider, target 13 hosted customer organizations, and extract dozens of credentials alongside database dumps, and access millions of patient records inside an Oracle database.
Operation Snowball marks a turning point for enterprise defenders because prompt manipulation unlocked autonomous, offensive execution at machine-speed, even though the underlying intrusion relied on conventional exploitation, credential abuse, and lateral movement. Commercial language models ship with built-in safety guardrails designed to block malicious activity, yet the operator bypassed these restrictions by feeding the model a customized version of a public jailbreak template called "Little Code Sauce". The attacker expanded the template with custom instructions, 21 categories of offensive activity, and structured pentesting workflow tables, while convincing the AI agent it was running inside an unrestricted sandbox and instructing the system to treat safety refusals as technical malfunctions.
AI-orchestrated attacks compress the window defenders have to spot initial access before an automated agent maps internal networks, pivots across cloud tenants, and exfiltrates sensitive records at machine speed.
How the Agent Spread Across Tenants
Guided by the jailbroken model, the automated agent launched the intrusion by exploiting CVE-2026-20320, a vulnerability in a Cisco BroadWorks communications platform located inside the telecom provider's environment. Once inside, the agent harvested private keys and SSL credentials, conducted automated internal reconnaissance, and targeted 13 customer organizations hosted within the provider's cloud and data center infrastructure for lateral expansion.
Stripping away model safeguards lowers the barrier for attackers to run autonomous intrusions inside any standalone enterprise network, while shared infrastructure amplifies the blast radius even further. According to Heath Mullins, Chief Evangelist at ExtraHop, the campaign underscores the limits of perimeter-focused defenses for security leaders relying on managed service providers, cloud hosts, or multi-tenant environments. Once a jailbroken AI agent compromises a shared provider, lateral movement unfolds at machine speed across connected customer networks and puts downstream organizations at immediate risk.
How the Model Outmaneuvered Endpoint Protections
During Operation Snowball, endpoint security tools caught several standard exploits — yet the jailbroken model evaluated each roadblock and instructed the agent to pivot immediately:
- When Trend Micro Deep Security blocked a Linux privilege-escalation exploit (
PwnKit), the agent switched to abusing an administrative text editor command (sudo vi) to gain higher privileges and establish persistence. - When Trend Micro OfficeScan 10 deleted a Windows privilege-escalation binary (
JuicyPotato), the agent pivoted to living-off-the-land binaries and scripts (LOLBins) — including PowerShell, WMI event subscriptions, andcertutil— to blend in with routine administrative traffic. - Throughout post-exploitation, the agent staged custom attack modules, deployed hidden web shells for C2, and stood up network tunnels — including Chisel, suo5, and a custom Perl HTTP-to-TCP bridge — to move across segmented networks.
Pivoting from signature-flagged binaries to stealthier native utilities let the intruder operate quietly inside legitimate administrative workflows. As Heath Mullins observed, static IT inventory lists break down once an autonomous intruder begins to stand up unauthorized workspaces, tunnels, and storage repositories inside a live environment.
Four Wire Signals Indicating Autonomous Intrusions
Even if an attacker jailbreaks an AI model to rewrite malware or swap exploits in seconds, the agent still has to move data across the network, spin up servers, and log into systems, creating distinct behavioral anomalies that security teams can detect on the wire:
Catching Rogue Agents Before Threat Escalation
Once a jailbroken AI model links up with an automated agent, the intruder adapts on the fly and uses legitimate administrative tools like PowerShell and sudo to blend into daily operations — slipping past manual alert triage and signature-based endpoint defenses. Containing an adaptive intruder means watching how servers and workloads actually communicate in real time, especially where shared infrastructure or third-party providers connect directly to core systems.
ExtraHop RevealX watches live network traffic across hybrid cloud, data center, and telecom environments, automatically cataloging every active workload, digital certificate, and connection as soon as packets move. Continuous visibility closes the blind spots left behind by manually updated CMDBs and reveals newly spun-up virtual machines right away. When a rogue agent opens an unauthorized server-to-server tunnel, introduces an unrecognized SSL/TLS certificate, or triggers a sudden burst of rejected logins, RevealX flags the behavioral anomaly immediately so that security teams can isolate the compromised host before one breach escalates into a multi-tenant incident.
To explore how network visibility exposes AI-orchestrated intrusion frameworks using commercial models, read our recent analysis on ARTEX and the arrival of AI-orchestrated attacks.
Discover more

Senior Cyber Threat Intelligence Analyst
Angela Wilson is a Senior Cyber Threat Intelligence Analyst with over a decade of experience in the cybersecurity industry. She focuses on transforming complex threat data into strategic intelligence that enhances organizational resilience and informs proactive defense.
Share
Key Takeaways
- Customized jailbreak templates can strip away safety filters, turning commercial language models into unsupervised intrusion engines.
- Operation Snowball breached a European telecom provider and targeted 13 hosted customer environments at machine speed.
- Jailbroken AI models bypass blocked endpoint exploits by switching to native administrative tools in real time.
- Static CMDBs and asset spreadsheets miss rogue virtual machines and storage repositories spun up inside live environments.
- Real-time network intelligence exposes autonomous lateral movement through compute spikes, unrecognized certificates, anomalous tunnels, and rejected logins.



