Blog
Notes from the boundary.
Research notes, field observations and long-form writing on agent security, LLM red teaming and MCP boundary testing.
Posts
-
The attack surface most teams ignore
Prompt injection is one door into an agent. What the agent is allowed to do, what it remembers and what it trusts usually matter more.
-
Three questions a security lead asks about your AI agent
Security reviews don't ask how accurate your model is. They ask what the agent can do, where the logs are, and whether anyone has tried to break it.
-
What an adversarial test report should look like
A scan report doesn't prove your AI agent holds up. Here is what one finding in an adversarial test report should hold, so you know what to ask for.
-
Your agent's tool surface is your attack surface
MCP lets an agent call tools with arguments it chose from context you may not control. Each tool deserves the scrutiny you give a public API endpoint.
-
Indirect prompt injection is not a user input problem
Most teams test whether users can trick their AI agent. The real attack is in the data the agent reads, not the prompts users type.
-
The SOC 2 gap your AI agent falls through
Your SOC 2 pentest covers your web application and your API. It does not cover your AI agent's tool access, credential exposure, or runtime authority.
-
Why your MCP server needs a red team
MCP servers expose tool descriptions, file access and shell execution to LLM agents. A unit test cannot find what an adversary will.
Subscribe
Seen enough to talk?
A few questions, one business day for a reply, and a real person on the other end of it.