TECHNICALAug 5, 2026 · 5 min read
When Near Zero Fails the Real Test
Language model safeguards that look nearly unbreakable on fixed tests can fail once attackers adapt, changing what release approval must require.
Alex Georges, PhDRead →
Research & Blog
PhD-level technical analysis of papers, benchmarks, and attack methods.
Language model safeguards that look nearly unbreakable on fixed tests can fail once attackers adapt, changing what release approval must require.
A release manifest, configuration comparison, and observed use review can keep enterprise approval tied to the AI agent that is actually running.
AI risk officers who approve assistants on text prompt results risk blocked procurement and lost revenue when combined images or documents expose a policy failure.
A formally correct extension of Gödel to AI guardrails tells attackers and defenders nothing they didn't already know. Its real payload is what it does to approval paper trails, and the math worth studying is in how guardrail failure actually scales.
How abstract topological data analysis research from grad school became the foundation for detecting AI hallucinations at AetherLab.
A technical deep dive into detecting and preventing hallucinations in large language models, from semantic checks and confidence scoring to adversarial testing.
A deep dive into keeping multi-agent AI systems reliable: orchestration, validation gates, conflict resolution, and monitoring for when AI agents work in teams.
An 8B parameter model just matched a 70B giant while running 100 times cheaper. Why smaller, sharper models are the future of practical AI deployment.
Occasional notes on AI risk, adversarial testing, and governance. No marketing drip.