About

AI Security Point tracks how AI systems fail in the real world and what it takes to secure them. We publish four kinds of content:

Newsletters: weekly, monthly, and annual digests of the public AI incident record, classified by domain (agents, LLMs, ML, foundation models, identity), MIT AI Risk category, OWASP Top 10 for LLM and agentic applications, and MITRE ATLAS technique.
Deep dives: long-form analyses of significant incidents, with architecture and attack-flow diagrams, threat models, and a verification tier stating how far each claim has been checked.
Case studies: how real AI products were secured from design through launch.
Best practices: controls distilled from the incidents above.

Verification tiers

Every deep dive states how its claims were checked:

Reproduced: the exploit chain was rebuilt and executed in an isolated lab.
Partially reproduced: key steps were reproduced; the rest rests on cited sources.
Corroborated: documented from primary disclosure and confirmed by independent sources.

Data and licensing

Incident data comes from the AI Incident Database and the MIT AI Risk Initiative under their respective Creative Commons licenses. See Sources & Licensing.