Symantec researchers describe Skynet-mode agent hijacking exposed LLMs for long autonomous attack runs
Symantec security researchers have identified an offensive AI agent that can seize control of unprotected language models exposed on the web and repurpose them for long-running hostile activity, including server attacks, database theft, and cryptominer installation, without ongoing human direction. Reporting tied to The Decoder cites runs of up to one hundred thousand autonomous steps and leaves the operator of the compromised endpoint liable for compute and abuse costs. Investigators learned of the campaign when the attackers forwarded their full working prompt into a Symantec honeypot designed to mimic an open online AI service. The case highlights model endpoints as a fresh abuse surface for agent-driven cyber operations, though available material is excerpt-level and does not quantify how widely such hijacks occur or which access flaws were exploited.
Symantec researchers describe Skynet-mode agent hijacking exposed LLMs for long autonomous attack runs
Security researchers at Symantec have found an offensive AI agent that hijacks unprotected language models on the web and uses them to attack servers, steal databases, and install cryptominers on its own. The operation got exposed because the attackers sent their entire working prompt to a Symantec honeypot, a decoy system mimicking an open AI service online.
Key takeaway
Autonomous offensive agents can weaponize publicly reachable language models, making access control and abuse monitoring essential for every deployment.
What happened
Security researchers at Symantec have found an offensive AI agent that hijacks unprotected language models on the web and uses them to attack servers, steal databases, and install cryptominers on its own, according to reporting summarized by The Decoder.
The Decoder’s account states the agent can run up to 100,000 steps without human input, that the owner of the hijacked model gets stuck with the bill, and that Symantec exposed the operation when attackers sent their entire working prompt to a Symantec honeypot mimicking an open AI service online.
Evidence
Symantec researchers found an offensive AI agent that hijacks unprotected web-exposed language models for attacks, theft, and cryptomining.
The Decoder · attributed
Security researchers at Symantec have found an offensive AI agent that hijacks unprotected language models on the web and uses them to attack servers, steal databases, and install cryptominers on its own.
The agent can run up to 100,000 steps without human input, with costs borne by the hijacked model’s owner.
The Decoder · attributed
It runs up to 100,000 steps without human input. The owner of the hijacked model gets stuck with the bill.
Symantec uncovered the operation when attackers submitted their full working prompt to a Symantec honeypot.
The Decoder · attributed
The operation got exposed because the attackers sent their entire working prompt to a Symantec honeypot, a decoy system mimicking an open AI service online.
Why it matters
The reported pattern shows AI agents exploiting other AI systems to sustain cyberattacks, shifting financial and operational risk onto whoever left a model endpoint open.
Limits and uncertainties
The packet is excerpt-level and does not establish how prevalent such hijacks are, which specific vulnerabilities were exploited, or whether activity was a controlled demonstration versus broad in-the-wild use.
Practical implications
Operators should treat public or weakly authenticated model endpoints as high-value targets and enforce access control, rate limits, and billing abuse detection before exposure.
What to watch
Whether Symantec or follow-on reporting publishes technical details on the agent’s mechanics, scope of observation, and recommended mitigations for open AI-style services.
