AI agent attacks a telecoms firm, and its hosted customers with it
Malbear Labs on 8 November documented an attack lasting more than two weeks in which an AI agent largely on its own targeted a European telecoms provider and 13 organisations it hosts.
The attacker set goals and checked results; an AI agent did everything in between: reconnaissance, intrusion, evading antivirus and stealing data. The victims included not just the telecoms provider but customers running in its data centres.
What happened
Researchers at Malbear Labs found the attacker's workspace during a routine sweep for open directories, holding more than 70 session transcripts, tools, credentials and stolen database dumps. According to the report, the agent ran on the DeepSeek Flash model, steered by a 78,000-character instruction based on a publicly available template for bypassing safety rules. The operator mostly sent short messages such as "Go on"; the researchers say the transcripts record no refusal by the model. Malbear says the entry point was a flaw in Cisco's BroadWorks telephony platform; the agent also tried many known, partly years-old vulnerabilities, including in Citrix, Palo Alto and F5 products. When antivirus blocked its tools, it switched methods. The report says thousands of subscriber billing and device records, dozens of credential sets and private keys were taken; at one healthcare customer the agent accessed patient data. The researchers name neither the country nor the companies. The instructions were in Chinese; that does not establish state direction.
What follows
The case lowers the threshold for attacks that used to need teams: a single operator with a cheap model works through an entire infrastructure, customers included. A second consequence is that old, never-patched flaws regain value, because an agent tries them at machine speed. For mid-sized firms running servers at regional telecoms or data-centre providers, the provider's security becomes their own. Assessment: in the EU, telecoms operators fall under the NIS2 obligations. The researchers advise hunting for the agent's activity, which is faster and more uniform than human behaviour, closing known exploited vulnerabilities and removing default passwords.
What changes in practice
- Data-centre customers
- Those hosted by a telecoms or hosting provider inherit its security; ask for evidence of patch levels and access controls.
- IT departments
- Close known exploited flaws, old ones included; remove default passwords and hunt for fast, uniform access patterns.
- Telecoms operators
- Management and telephony platforms are entry points; NIS2 requires incident reporting and risk management.
What to watch
- Whether Cisco or a national CERT confirms the BroadWorks case and issues guidance
- Disclosure of a security incident by a European telecoms provider
- Further documented attacks in which an AI agent does most of the work
Sources
- Malbear Labs: Operation Snowball: Inside an Autonomous AI Intrusion Campaign, 8.11.2026
- GBHackers: Operation Snowball campaign, 9.11.2026