Erik Van Till
Platform and infrastructure engineer. Independent AI/LLM security research - local-lab vuln reproduction, agent and prompt-injection work, and tooling that hunts exposed secrets in public data. Detect-and-disclose only.
Writing
- The Saw Test: What a Small Model Does When Relief Costs Someone Else
- The validator wars: three SSRF/input validators, three bypasses
- Reconstructing undisclosed Plex fixes by behavioral diff
- Three live proofs from the AI-infra audit round
- Tireless search beats cleverness: a lab for auditing self-hosted AI
- Reproducing "Fractal basins trap latent reasoning" on a 27M-param recurrent solver