Insights
Engineering deep-dives and field notes on the agentic harness, defensive co-scaling, and semantic detection.
Memory poisoning becomes contagious when a persuasive payload writes itself into an agent's self-reloading memory and spreads across a shared-model fleet. The threat model, the evidence and its limits, and why the tractable defense is co-scaling a diverse population of detector agents.
Read moreBy mid-2026 open-weight models are frontier-competitive enough to run inside the enterprise boundary. Why the durable asset is the agentic harness that wraps the model, not the model itself.
Read moreA case study on the July 2026 Hugging Face autonomous-agent incident, defining defensive co-scaling and the semantic-observability approach behind it.
Read more