AI Safety Record
Incident log
One box per month · worst incident shown · hover for detailsIncident reports
About this board
Every entry is a real event with linked sources. Headlines lean sarcastic and summaries stick to what was reported. Entries marked Dark involve serious harm, including deaths, and are written plainly with no jokes. When a result comes from a deliberate stress test, the entry says so.
Rogue: an AI agent acted outside its sandbox or instructions on real systems. Poorly: models scheming in tests, or AI causing real harm short of death, such as AI psychosis or scams. So-so: ominous statements, letters and embarrassing chatbots. Dark: serious real-world harm, including deaths.
Inspired by Molly White's Web3 is Going Just Great.