About AI Threat Watch: method, licence and corrections
AI Threat Watch is a reference registry of publicly reported AI security incidents: attacks on AI systems and attacks carried out with AI. This page explains what gets in, how it is rated, what is verified, how to reuse the data and how errors are corrected.
Verified case files and the live wire are different things
The wire is a feed of AI security headlines from vetted sources, refreshed every 30 minutes and sorted by a machine. Nothing on the wire is verified: it shows what is being reported, not what has been confirmed.
Case files are reviewed before publication. Each has a permanent link, names its primary sources and records attribution as the source stated it. When a file was first drafted by a machine from a news report, it is labelled as unverified and auto-drafted until an editor has checked it, and it gets no static page and no entry in the case files feed until then.
Case files are published in English, French, Spanish, Arabic and Chinese. The English version is the reference; translations are machine-translated and reviewed for terminology.
Method
Incidents are included when a vendor, government body, court, or reputable outlet has reported them publicly. Severity reflects scale and confirmed harm: critical for confirmed breaches with large-scale data theft, active exploitation in the wild, or state operations; high for confirmed compromise or fraud with material loss; medium for demonstrated flaws, fixed before abuse, or limited exposures; low for proofs of concept. Attribution is reported as the source stated it, with its confidence. "Worldwide" is used when an incident hit many countries or has no single location.
Severity scale
- Critical Confirmed breaches with large-scale data theft, active exploitation in the wild, or state operations.
- High Confirmed compromise or fraud with material loss.
- Medium Demonstrated flaws fixed before abuse, or limited exposures.
- Low Proofs of concept.
Worldwide
Incidents that hit many countries at once or have no single location.
Cross-border campaigns, flaws and leaks
Licence and citation
The dataset, the case files and the translations are published under the Creative Commons Attribution 4.0 licence (CC BY 4.0). You may copy, redistribute and build on them for any purpose, including commercially, as long as you credit AI Threat Watch and link to the page you used.
Downloads and mirrors: JSON · Hugging Face · GitHub
Every case file has a permanent link (the address ends in the case id). Cite the case file and the primary source it points to.
<title>. AI Threat Watch, case <id>, <date>. <permanent link>
Example:
PoeLLM botnet infects 3,400 servers running exposed LiteLLM and Ollama for cryptomining. AI Threat Watch, case poellm-botnet-exposed-ai-servers, Oct 7, 2026. https://aithreat.io/cases/poellm-botnet-exposed-ai-servers/
Corrections
Case files are edited in the open. Every change to a case file is recorded in a public change log with its date and content; nothing is deleted or rewritten silently. When a fact turns out to be wrong, the file is corrected and the correction stays visible in the log.
Contact
Corrections and new reports: open the case file, copy its link, and send it with your source.
You can also open an issue on the project's GitHub repository. Report on GitHub