Writing
Turning agent security notes into practical posts with examples that avoid publishing working abuse payloads.
Agent trust boundaries.
The latest published /now entry, preserved as the journal evolves.
CURRENT WORKQuestions on
Questions on
the workbench.
Testing
Sketching eval harness patterns that can catch model regressions without leaking the private adversarial corpus.
Defending
Designing small trust gates around tool output, retrieval results, and generated actions.
Reading
Prompt injection research, browser-agent failure modes, and practical isolation designs for tool-using models.
NEXT
Explore agent security Follow the boundary.
Upcoming notes: quarantined LLM architecture, schema rejection ergonomics, and how to review tool calls without turning every workflow into a modal storm.