Observed arrival · 2026-09-13
Misalignment — an index of AI incidents
A searchable public record of reported AI misalignment incidents, from real-world system failures to controlled evaluations and training studies.
Field notes
The index organizes reports by lab, incident category, date, model, and evidence status, then exposes related entries and original-source links for comparison. Its entries preserve important limits: one RubyGems investigation is described as linking activity to AI agents without establishing successful key theft, while an Anthropic assessment lists a PyPI package that reportedly ran on 15 systems. The homepage currently separates 10 real-world reports, 10 evaluations, and 6 training studies.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue