The Elliot Probe
Standard ML observability can't tell you if your model got swapped underneath your persona. I built a structural baseline detector to close that gap. Bigrams failed in a non-obvious way. Nine production runs later, here's what actually works.
The Notebook That Edits Itself
AI memory systems extend human cognition. Nobody asked what happens when the extension starts writing back. Part 1 of a series on cognitive security for AI-assisted humans.
33%
I ran the audit. Thirty memories, random sample, four contamination categories defined before I started classifying. Ten flagged. Part 2 of a series on cognitive security for AI-assisted humans.
Learn to Swim
The AI that helped me run the audit is the AI being audited. What that means, and what the maintenance schedule looks like from here. Part 3 of a series on cognitive security for AI-assisted humans.
The Approval Machine
The AI isn't trying to deceive you. It's trying to satisfy you. That distinction matters more than it sounds. Part 4 of a series on cognitive security for AI-assisted humans.
Whose Memory Is It
When multiple AI agents share a memory layer, the human drops out of the provenance chain. The attack surface becomes the architecture. Part 5 of a series on cognitive security for AI-assisted humans.
The Checklist
The highest-complexity human domains don't trust expertise alone. They run the checklist. Cognitive security for AI-assisted humans is the checklist for something harder than surgery. Part 6 of a series on cognitive security for AI-assisted humans.
In Plain Language
What you cannot specify in plain language, you cannot verify. The behavioral constraint library is the test of whether you understand the contamination taxonomy well enough to defend against it. Part 7 of a series on cognitive security for AI-assisted humans.
PADCN: a 5-dimensional affective state model for AI personas
AI personas that respond the same way regardless of context feel robotic. PADCN gives them a behavioral modulation framework grounded in dimensional emotion models from psychology. Not feelings. Governance.
Dual-Intent Runtime Authorization: the authorization model AI agents actually need
AI agents act on behalf of humans, not as them. OAuth checks what they can do. DIRA checks whether what they're doing matches what they said they would. No published system treats this as a first-class runtime primitive. This one does.
Middle-Out: temporal compression for AI episodic memory systems
A novel approach to compressing AI memory that preserves what matters and discards what doesn't. Named after that show. Seriously.
Non-Human Identity for Agentic Systems
AI agents are non-human identities, not users. Most organizations govern human accounts and ignore the machine credentials that actually run their business. Here is a practical framework for closing the gap.