AI Security

There are two sides to AI security, and most teams are only looking at one. There is securing the AI systems you build: models, agents, tools and the protocols that connect them. And there is the security of the software AI now helps you write. I write about both, and the thread running through all of it is the same: establish security properties at design time, because a probabilistic model cannot give you a deterministic guarantee at runtime.

Securing the AI systems you build

• Threat Modeling AI Systems: A Complete Playbook for Practitioners: the adversarial subspace, transferable decision boundaries, and a full kit for running sessions, including the Elevation of Autonomy deck.
• Threat Modeling MCP: why the Model Context Protocol trust model bites, the lethal trifecta, and deterministic controls first.
• Elevation of Autonomy on GitHub: 30 cards in five suits for AI, LLM and agentic systems.

AI in the software development lifecycle

• The Threat Model Was Never the Point: a train-the-trainer kit for AI-assisted threat modeling. What happens with thin context, a weak skill or a cheaper model, why the room still matters, and a repo to run it yourself.
• Is Your AI Philosophy Broken?: the productivity gains of AI-generated code against its security cost.
• Cursor as your Secure Dev Team: using an AI coding tool to support secure development.
• Software Engineering Has Regressed 50+ Years Since AI: what we gave up in the rush.

AI and static analysis

• Why SAST is Broken!: where static application security testing (SAST) falls short.
• SAST vs Claude Code Security: A Deep Dive: traditional tooling against an LLM-based reviewer.
• src2sink: why LLM-based SAST needs deterministic ground truth, and a tool that provides it.

Where to spend the effort

• Sparse or Dense? Claude Mythos, Dan Geer, and Where We Should Actually Be Spending Our Effort
• Get PQC Ready PDQ, Part 1: the post-quantum migration window and why AI shortens it.

Note to architects: if you read only one post here, read the playbook. Every other post on this page comes back to its central argument.

Stay tuned!