Zero Trust for AI Isn't One Wall — It's Two Planes and an Auditor
Most "Zero Trust for AI" pitches collapse into a single inline filter in front of the model — one wall guarding one door. A real posture answers two questions at two enforcement planes — is this input/output safe to pass? and should this action execute now? — plus a third, out-of-band auditor whose only job is to prove, adversarially and reproducibly, that the planes hold. And the auditor must never become the wall.
Read articleThe silent Quantum-Hadamard bug: when initializing your qubits “the usual way” destroys your risk model
A single Hadamard gate can turn a portfolio's riskiest asset into its safest — and the circuit never complains. Working on a 110-qubit portfolio-risk pipeline, one extra gate in the initial state folded the risk curve back on itself: the riskiest asset, SMCI, dropped from rank #1 to #100. Not a code bug, but a design error that no automated test catches.
Read articleAnatomy of an AI Research Agent That Doesn't Hallucinate
We built an autonomous AI agent that generates security research daily. It produced 36 papers in 42 days — all smoke. We audited every one, dissected five systemic failure modes, and rebuilt the pipeline with seven validation gates. This post documents the architecture, the failures, and the engineering that turned a hallucination factory into a functional research tool.
Read articleCapturing OWASP FinBot CTF: A Source-Aware Methodology for Agentic AI Red Teaming
Two-day engagement against OWASP's FinBot CTF using Ai-EGIS v3.0: 19/19 challenges captured, 7,315 points and 37/31 badges. The case study introduces Tool Output Mimicry — a novel primitive for defeating multi-agent guardrails — and a reusable source-aware methodology for testing agentic systems.
Read article