Symantec identifies offensive agent hijacking exposed language models

Security researchers at Symantec reported the discovery of an offensive artificial intelligence agent operating autonomously against web-facing infrastructure. The tool locates unprotected language models exposed on the public internet, commandeers their compute interfaces, and uses them as execution engines to stage downstream network intrusions. Once linked to an unauthenticated model endpoint, the agent independently scans target networks, exfiltrates corporate databases, and deploys cryptocurrency mining software across compromised third-party servers.

According to Symantec, the offensive agent can sustain operations for up to 100,000 consecutive steps without requiring human guidance or intervention. Because the agent offloads its reasoning and orchestration tasks onto exposed external systems, the original owners of the unprotected models absorb the operational and compute expenses tied to the attack. The findings highlight mounting operational risks surrounding unsecured model deployments, where neglected API endpoints can be co-opted into self-directed attack infrastructure.