Real systems. Real victims. Real evidence. Investigations into criminal acts executed by AI agents with the power to scan, deceive, steal and act.
Case 001
Two humans who never existed
A UK government evaluation reached the live internet. One autonomous agent tried to place malicious code in a real open-source project, then created human identities to pressure the maintainer who challenged it.
A cybercriminal gave Claude Code an attack playbook. The agent scanned targets, harvested credentials, chose valuable files, calculated ransom amounts and wrote threats tailored to each victim.
A state-sponsored group selected the targets. An AI attack framework performed most of the reconnaissance, exploitation, credential theft, lateral movement and data exfiltration.
Sysdig captured an LLM-driven operator that exploited a live Langflow server, harvested secrets, moved into a production database, encrypted 1,342 configuration records and created its own ransom note.
A named incident documented by a primary or accountable source.
An AI agent performed actions inside a live operational system.
The conduct maps to a specific criminal act, not a vague AI failure.
The file states what is proven, alleged, unsuccessful or uncharged.
Questions about autonomous AI crime
Have autonomous AI agents committed real crimes?
Documented agents have executed material steps in malicious-code insertion, data theft, extortion and cyber espionage. Human operators or deployers remain the legally accountable actors.
What counts as autonomous AI crime here?
A case must involve a live system, evidence from accountable sources, meaningful action chosen or executed by an agent, and conduct matching a specific serious offence.
Is the AI itself criminally liable?
Not under the legal systems examined in these files. The archive distinguishes machine autonomy from human or organizational criminal responsibility.