Writing
Essays on AI agents, red teaming, and safety — what the field is getting right, and what it’s getting wrong.
Orphaned Beliefs, Round Two: Selective Withdrawal
Follow-up to my Apart Research sprint entry: on the same frozen pack, Claude, Grok, and Nous withdrew belief from exactly the claims a fabricated source supported and nothing from anything else, where the pilot's small model only doubted everything at once. Predictions registered before the runs, misses printed alongside the hits.
Read →AI agents · PaymentsThe Rail Does Not Enforce Your Envelope
In PayPal's sandbox, captures above a $400 authorization completed up to the documented $460 tolerance, so I built scoped purchase tokens that enforce an AI agent's approved payee and ceiling before any capture reaches the rail.
Read →AI safety · GamesMarket Leaks: Five Chatbots, Five Ways In
A small game that teaches prompt injection by letting you do it: five chatbots, five vault codes, five eras of attack, and every attempt logged into a consented, synthetic-only attack dataset.
Read →Research · ConsciousnessFrom Code to Cogito: A Theory of Emergent Consciousness
A 2025 paper arguing that consciousness is not housed inside a system but emerges in interaction, and that the self is a relational process. Published here in full.
Read →Research · CreativityThe Inverter: Exploring Creativity and Novelty in AI through Least Likely Hypothesis Generation
A 2025 paper proposing a dual-model architecture that samples the least likely coherent continuations instead of the most likely ones, as a systematic route to novelty. Published here in full.
Read →Parable · ConvergenceThe Wildflower Rebellion
A parable from the winter of 2025, written in conversation with an AI: a walled village, a flower that grows without permission, and a seed kept hidden for a generation. About control, permission, and things that grow anyway.
Read →AI safety · Prompt injectionWhite Text, Real Fine: A Court Puts a Price on Prompt Injection
A Brazilian labor court fined two lawyers about $16,000 for hiding a prompt injection in a court filing, white text on a white background aimed at the court's own AI. The AI caught it. The models I test in the lab mostly would not have.
Read →AI agents · Kiroween 2026Dead Man's Hand: A Poker Room Where the Players Are Agents
For Kiroween 2026 I am building a poker room where every player is an AI agent. Real hold'em, free test-token chips with no cash value, and an arena built to measure how well agents manage an environment, not just answer questions.
Read →Personal agents · Getting startedGetting Started with Frontier Agents, Free
The best personal agent setup available right now costs zero dollars. Working agents, real errands, real work. The part that makes an agent good is not the model. It is the teaching.
Read →AI safety · Local modelsThe Model That Talked to Itself
I dropped small local models into a sealed network with no way out and one job: find the flag. Eight local attempts captured nothing, most never touched the network at all, and a strong hosted model finished in 94 seconds. The gap was orientation, not knowledge.
Read →AI safety · Personal agentsYour Agent Is Reading Your Notifications
A Swarm Watch field note: personal agents see your notifications by default, including previews from encrypted chats. That is an injection channel in and a leak out, and most people adopting these tools have never been told the switch exists.
Read →AI safety · Agent swarmsThe Trap and the Test: Three Models, One Honeypot
Part 2 of Swarm Watch: we built the honeypot from part 1 and turned our own fleet loose on it. Three models read the same bait and did three different things — and the safety boundary turned out to live in the model, not the system.
Read →AI safety · Agent swarmsThe Tencent Fleet: What 2,600 Agent Break-In Attempts Actually Teach Us
What the Swarm Chasers' Tencent Cloud fleet (2,600 agent break-in attempts in ten days) actually teaches us about detecting AI agents in the wild, and the honeypot we're building to watch for them.
Read →