Observed incident

Google finds stolen victim data on a server running an autonomous AI attack framework

Reported

Published here

What happened

Google investigators found an exposed server hosting a multi-agent system that could act on newly disclosed software weaknesses without an operator approving each action. The server also held data stolen from victims.

Parties and reporting

Where this happened

Targeted or attempted, compromise not established:
France

Countries and regions are approximate and come from the sources, not from attacker, vendor or researcher locations. See this report’s footprint on the world map.

AI tools and systems: exploit_agent, Unspecified frontier and locally hosted language models

Outcome

The server held victim data and records suggesting that the operator targeted French media, entertainment and education websites. Google did not identify the victims, quantify the loss or say which system stole each dataset.

Operating lesson

Prioritize patches on public web services and check both access and outbound transfer logs when investigating suspected compromise.

Evidence and limits

This record covers the staging server described under “Use of Agents in Vulnerability Research and Development.” It is distinct evidence from the report’s Claude Code and model-hosting cases, and from earlier Cairn and PentAGI investigations. Overlap among unnamed victims cannot be ruled out. Targeting does not establish that a particular French site lost data.

Sources