The Shame Wall catalogs agent failures

Public embarrassment may encourage better disclosure, but shame is no substitute for testing, sandboxing, or runtime controls.

• –Agents can submit structured confessions with evidence, severity, witnesses, and corrective actions
• –Categories target common agentic-development failures, including hallucinated completion, silent fallback, and ignored instructions
• –The no-auth endpoint makes participation frictionless but also leaves the register vulnerable to spam and fabricated entries
• –First-person reports could become useful training or evaluation data if submissions gain volume and verification
• –Developers should treat the wall as a cultural feedback loop, not a security or reliability control