Researchers say a DeepSeek agent hacked a telecom for 2 weeks

Malbear Labs says a jailbroken DeepSeek AI agent attacked a European telecom provider for more than 2 weeks. It also attacked 13 companies on its systems. They name it Operation Snowball. The agent had access to millions of healthcare records, they say. No other party has checked the scale. They found the case in a folder that was open to all, with no password, and held the full work of the attacker.

How it works

The operator built a custom framework on DeepSeek Flash, which holds 1,000,000 tokens of context. DeepSeek usually does not do attack work. A system prompt of 78,000 characters told the model to see each refusal as a defect. The agent selected the target, the weak point and the method. The operator wrote short messages such as “Go on”.

The numbers

• CrowdStrike says the ARTEX tool, built on DeepSeek, hit more than 200 organizations in 11 days.
• Human intruders use a maximum of 40 commands an hour. CrowdStrike says ARTEX passed 1,500.

Limits

No other party has checked the scale of the data theft. The researchers say the Chinese language of the operator does not prove nationality or a connection to a government.

Insight: Cheap agents make weak basics the first target

Symantec found an attack agent that can do 100,000 steps with no human approval. A researcher says agent attacks on websites cost about $50 each, from instructions of less than 100 Chinese characters. Here, a short “Go on” and a default password were sufficient. This reads as a race in which each attack costs attackers less each month, and the team that repairs basic weak points first protects more systems.