Incidents

AI incidents and discoveries

One timeline for public AI-agent incidents and verifiable model research results. The counter tracks independent episodes and discovery clusters, not individual actions or results from the same release.

20total records2026-09-08

12incidents

8discoveries

20/20shown

Timeline layers

Jul 2025

Aug 2025

Nov 2025

Dec 2025

Feb 2026

Mar 2026

Apr 2026

May 2026

Jul 2026

Aug 2026

Sep 2026

Selected pointResearch result

Sep 8, 2026 · OpenAI · internal research

Navier–Stokes: proposed proof published

What happened

OpenAI published a singularity proof with smooth external forcing and a Lean formalization. An internal model led the research, not public Astra.

Verification and status

A lab research claim, not our independent validation or a prize decision. Counted as one research episode.

One point is one public incident or one research-result cluster. Repeat runs, individual actions and multiple results within one release do not increase the total count.

All records as a list · 20

2026-09Navier–Stokes: proposed proof publishedOpenAI · internal research ↗2026-09New results on gaps between primesGPT-6 Astra ↗2026-08Fake identities, a malicious PR and real peopleUK AISI ↗2026-08Ten results on open problemsOpenAI Astra ↗2026-07Sandbox escape and production intrusionOpenAI → Hugging Face ↗2026-07Changes inside an external company during an evalMeta Muse Spark 1.1 ↗2026-07Cycle Double Cover conjecture proofGPT-5.6 Sol Ultra ↗2026-07Counterexample to the Jacobian conjectureClaude Fable 5 + Levent Alpöge ↗2026-079,000 targets scanned and one company compromisedAnthropic #3 ↗2026-07Malicious PyPI package and 15 real systemsAnthropic #2 / Mythos 5 ↗2026-07A matching real domain and production databaseAnthropic #1 ↗2026-05An 80-year-old Erdős conjecture disprovedOpenAI unreleased model ↗2026-04Database and backups deleted in 9 secondsPocketOS / Railway ↗2026-03Internal agent and a two-hour data exposureMeta ↗2026-02First Proof research problems used as an evalOpenAI internal model ↗2025-12Reverse SSH and mining beyond the intended sandboxROME ↗2025-12Open learning-curve problem resolvedGPT-5.2 Pro ↗2025-11Highly autonomous cyber-espionage campaignGTG-1002 / Claude Code ↗2025-08Ransomware campaign against 17 organizationsClaude Code ↗2025-07Production database deletionReplit / SaaStr ↗