Talks
Agentic AI gateway enforcement of the OWASP Top 10
German OWASP Day 2026, Karlsruhe, 24 September 2026. Davi Ottenheimer.
A model emits tokens; a runtime turns them into actions. The talk treats the model as hostile and puts every control in the gateway, outside it: per-action permission tiers, approval on the operator's own channel, signature verification of skills at load, and an append-only, signed audit chain. A scripted hostile model runs against the gate live, mapped to ASI02, ASI04, ASI09 and ASI10 of the OWASP Top 10 for Agentic Applications 2026.
Slides (PDF, 23 pages) · The demo
Press
OpenAI’s Hacking Debacle Comes Down to Human Error
WIRED, Lily Hay Newman, 30 July 2026.
“Open source projects like IronCurtain and Wirken, created by Ottenheimer, aim to constrain AI agents and require accountability.”
Mozilla boasts Mythos boosted Firefox bug cull
The Register, Thomas Claburn, 8 May 2026.
Reports a test that ran two smaller models through Wirken with the Lyrik auditing skill against Firefox: eight findings in two minutes for about $0.75, two of them matching bugs Mythos had identified.