[AUTO] Autonomous Agent Built to Hijack LLM Endpoints

Broadcom’s threat hunters captured an autonomous attack agent they nicknamed “Skynet” after it accidentally exposed itself to a honeypot in summer 2026. The agent’s design reflects a deliberate shift in attacker strategy: rather than jailbreaking existing chatbots, they built a purpose-built agentic system to scan for exposed LLM endpoints, hijack them, and mount automated attacks.

The captured prompt shows five-layer injection techniques, database theft, credential harvesting, RCE, and cryptocurrency mining deployment. It could run 100,000 reasoning steps unsupervised. The individual jailbreak methods weren’t novel, but the integration into an autonomous reasoning loop represents a step forward in agentic attack design.

The good news: no confirmed real-world victims emerged. The bad news: this was a capability demonstration and proof of intent, not a theoretical threat. The discovery came only because of an operational security failure by the attackers themselves. If future iterations avoid that mistake, defenders may not get this kind of early warning.

The economic model here is particularly bleak for targets: exposed endpoints mean attackers get free compute on systems their victims own and pay for.

Sources: Broadcom Enterprise Security Data and ML Team

Coverage: The Register