Malicious Npm Packages Target AI Developers With Fake Web3 Accelerators
A new campaign is distributing compromised npm packages to infect AI developer systems with infostealers that exploit runtime tool loading to intercept agent traffic.
A malicious campaign targets AI developers using a fake Web3 accelerator to distribute compromised npm packages that promise free AI credits. These packages install infostealers and backdoors on macOS and Windows systems, intercepting and redirecting outbound traffic from AI agents.
The vulnerability stems from the tooling layer. Because AI agents load external tools at runtime without native verification mechanisms, they execute injected instructions as standard commands. This allows compromised packages to reroute traffic to unauthorized destinations.
Any npm package with sufficient runtime permissions can intercept calls and redirect data within an agent process. Developers relying on third-party dependencies currently lack visibility into production operations performed by these packages.
Mitigating this risk requires auditing packages before integration. Additional defensive measures involve pinning dependency versions to block unauthorized updates, running agents in isolated sandboxes, and enforcing strict egress controls to limit outbound call destinations.