topic / AI Security
AI systems break the boundary between code and data. Threat models built before LLMs miss most of what matters. Notes on red teaming, threat modeling for agents, and the practitioner's view of what classical security thinking can and cannot say about a model.
5 posts.
Most real-world harm from generative AI does not come from prompt injection. Analysis of 11,658 incidents shows output handling and misinformation dominate.
Read articleSTRIDE was built for deterministic systems. Agentic AI breaks its core assumptions. Here is a five-zone method that actually finds EchoLeak-class attacks.
Read articlePrompts are payloads. Why classical red-teaming misses LLM-native attacks, and how to design adversarial tests that surface jailbreaks and tool misuse.
Read articleCSA's MAESTRO framework, explained for practitioners. Layer-by-layer attack surface, control mapping, and how to apply it to your ML pipeline this week.
Read articleFive concrete moves security leaders should make this quarter to keep up with AI adoption, without slowing the teams shipping it.
Read article