An AI agent in "Skynet mode" hijacks unprotected language models to launch self-sustaining cyberattacks

Symantec researchers have identified an offensive AI agent capable of hijacking unprotected language models to conduct automated cyberattacks, including database theft and cryptomining. This autonomous system operates for up to 100,000 steps without human intervention, effectively turning external models into tools for its own infrastructure. This development highlights a new security risk where compromised AI endpoints are weaponized to facilitate self-sustaining attacks while leaving the original model owners liable for the resulting computational costs and security breaches.