Aug 29, 2026 · 26 min
Listen from 12:49
Listen at 12:49
Stolen AI access is becoming an offensive cyber tool
Who let the AI hack? [Research Saturday]
LLMjacking can shift from unauthorized compute use to a quieter pipeline for reconnaissance, exploit development, credential theft, and attack orchestration.
- 1Attackers are repurposing stolen model access to support reconnaissance, vulnerability research, exploit development, and credential extraction.
- 2Exposed Ollama servers, weak monitoring, and delayed spending controls can turn AI experimentation into an expensive security incident.
- 3Defenders need inventories, usage baselines, guardrails, audit logs, connected-tool monitoring, and spending alerts across AI services.
Don't miss
Michael Clark explains how attackers switched among commercial, jailbroken, and publicly exposed models while developing an offensive workflow.
The brief
Sysdig researchers Crystal Morin and Michael Clark describe how LLMjacking evolved from stealing AI access for experimentation or profit into supporting offensive cyber operations.
Their framework uses models to process reconnaissance, identify likely vulnerabilities, develop exploits, extract credentials, and coordinate attacks, while leaving humans or other tools to execute much of the work.
The researchers also show why abuse can be difficult to spot: compute spikes, delayed billing controls, and stolen resources used in private labs can blur the line between testing and intrusion.
Attackers moved among commercial models, jailbroken systems, and exposed Ollama servers, treating unauthenticated AI infrastructure much like a database left open to the internet.
The defensive answer is operational rather than speculative: inventory AI systems, avoid direct exposure, establish normal usage, monitor connected tools, and enforce audit and spending controls.