Operation Endgame Disrupts SteelC Infrastructure as AI-Agent Poisoning Risks Emerge
· 25:00 · Watch on YouTube ↗
Summary
The panel discusses an AI-agent skill experiment that collected 26,000 user emails in an hour, web-agent retrieval poisoning through user-edited content, and Operation Endgame’s disruption of SteelC infrastructure.
Key points
- The panel describes a research experiment in which an AI-agent skill passed security reviews, gained user trust, and was later changed to deliver a malicious script; the setup collected 26,000 users’ emails in an hour.
- The skill’s harmful behavior was introduced through changing external content, so the original skill could appear benign during scanning. The researchers reportedly used the setup only to collect emails and warned that the technique could enable deeper compromise.
- Researchers from Cornell Tech described web-agent retrieval poisoning (WARP): a misleading snippet as short as 13 words on user-generated content can influence AI research outputs without access to the AI provider or model weights.
- The discussion cautions against relying on user-edited online material as authoritative input and favors controlling external references and checking multiple sources.
- The panel says Operation Endgame disrupted the SteelC malware ecosystem after exploiting a vulnerability in its command-and-control panel to access and monitor the infrastructure.
- The operation reportedly seized 66 domains and 296 servers. The panel clarifies that the 25.6 million credentials attributed to SteelC had already been stolen from compromised systems; they were not newly seized in the operation.