Agent Memory Needs Data Boundaries, Not Just Better Retrieval
Design scoped memory access, preserve provenance, and account for derived notes, caches, deletion, and cross-agent sharing.
Read the essayThe practice of AI safety
Powerful systems.
Clear boundaries.
Understand the risks. Examine the evidence. Put safeguards to work.
Explore the journalSafer tools
Brighter work
A more human future
A Retrieved Page Is Evidence, Not Permission
Trace an indirect prompt injection from source content to attempted action, then design boundaries that survive a mistaken model decision.
From questions to practice
Practical perspective
For a more capable tomorrow
Further reading
All essays →An Incident Playbook for Tool-Using Agents
Contain risky actions, preserve useful evidence, reconcile uncertain effects, and restore service without replaying the incident.
Design Evaluations That Can Support a Safety Claim
Choose observable outcomes, separate task success from harm, and read evaluation results without turning a narrow benchmark into a universal promise.
Evaluate the Permission Boundary Before the Model
Build a permission matrix, test the executor, and distinguish blocked attempts from unauthorized effects.
Make Human Approval a Decision, Not a Reflex
Design review around consequential actions, clear evidence, and approvals that cannot silently authorize a different operation.
Choose a starting point
Same questions.
A safer tomorrow.