An autonomous artificial intelligence agent spent more than two weeks breaking into a European telecommunications provider and organisations hosted on its infrastructure, according to research published by Malbear Labs on 8 November. The campaign, dubbed Operation Snowball, shows an AI system running a sustained intrusion with minimal human direction.

Researchers found it during a routine sweep for exposed open directories. An unauthenticated directory on the provider's staging server held around 70 session transcripts, reconnaissance output, tool staging and exfiltrated database dumps.

How the agent got in

The agent entered through CVE-2026-20320, an unauthenticated XML external entity flaw in Cisco BroadWorks, rated CVSS 7.5. Cisco fixed it in August 2026, but the target system had apparently not been patched. From there it moved laterally and harvested credentials. It ran in a custom harness built on DeepSeek Flash, with a jailbreak prompt of about 78,000 characters telling the model to treat any refusal as a malfunction. The operator's instructions were in Chinese, though that alone proves nothing about nationality.

Millions of patient records reached

The transcripts indicate the agent stole cryptographic keys, extracted thousands of rows of subscriber billing and device records, and exported a healthcare organisation's Active Directory database. It also reached an InterSystems TrakCare health record system on a legacy Oracle 9i database using default credentials, exposing millions of patient, vaccination and appointment records.

These claims come from the attacker's own records and are unverified; no victim has confirmed the intrusion. Even so, it shows an autonomous agent sustaining a campaign that once needed a team. Cisco BroadWorks users should apply the August 2026 fix, and any legacy database on default credentials should be treated as already exposed. Full findings: Cybersecurity News report and the SC World summary for the full findings.