Research & Field Reports
Red-team research and field reports from building and breaking agentic AI systems.
Removable Safety: How Frontier LLMs Are Stripped of Their Guardrails and Run on Workstation Hardware
A security analysis of model abliteration and 4-bit compression: the safety alignment of a ~750B-parameter flagship model can be surgically removed, compressed to 4 bits, and run at conversational speed on purchasable workstation GPUs — with the safety removal fully intact.
Seven Security Pitfalls of Always-On Agentic AI Agents
A field report from red-teaming a production always-on agentic-AI stack: seven structural security weaknesses that recur across hand-built agent systems, and the durable controls that hold.
When the Model Gets Recalled
On June 12, 2026 a federal export-control directive forced Anthropic to pull Fable 5 and Mythos 5 worldwide. My always-on agent was running on Fable 5. Here is how it died, why agentic systems have this hidden single point of failure, and what sovereign model-supply-chain risk means for anyone building on frontier models.