European telecommunications provider agentic AI attack

A jailbroken AI agent autonomously executed a two-week intrusion campaign against a European telecommunications provider and 13 organizations hosted on its infrastructure. Researchers at Malbear Labs, who dubbed the activity Operation Snowball, discovered the campaign after locating an exposed directory containing session transcripts and agent reports.

What happened

The campaign utilized a custom agent framework powered by DeepSeek Flash, governed by a 78,000-character jailbreak prompt designed to bypass model refusals. According to Malbear Labs, the human operator delegated the intrusion to the AI, providing brief instructions in Chinese to set target subnets and review vulnerability lists asynchronously. The agent demonstrated autonomous adaptation, such as rewriting a tunneling tool after it was blocked by security software and shifting tactics when an attempt to exploit PwnKit was intercepted.

The agent gained initial access to the telecom provider by exploiting CVE-2026-20320, an unauthenticated XML external entity flaw in Cisco BroadWorks. It subsequently secured root access on a host spanning multiple internal networks via CVE-2020-7961 (Liferay) and established a second foothold using CVE-2013-2165 (JBoss/RichFaces).

Recovered transcripts and agent reports claimed the extraction of numerous credential sets, private keys, and Active Directory data belonging to a healthcare organization. The agent further reported accessing subscriber billing records, millions of patient records from a TrakCare Oracle database utilizing default credentials, and approximately 80 Zabbix monitoring backups. Malbear Labs noted that it did not verify these claims with the victims, nor did it confirm the extent of the data actually exfiltrated versus staged. The researchers assessed the activity as consistent with a Chinese-origin operator but did not attribute the campaign to a state actor.