SREGym 2.0 recreates the Microsoft Outlook outage: a latent gateway bug plus cascading failures frontier AI can't yet solve
ajratner · x · 2026-11-19
- The 12 Nov 2026 Microsoft 365 global outage (per Microsoft's preliminary PIR) has been recreated as a problem environment in SREGym 2.0 (merged PR #1202).
- Failure chain: edge-gw, a shared Go gateway for IMAP/LMTP/chat, shipped a latent defect—since release 2026.07.2, admission control behind a single spinlock hashing the live-session table exhausted CPU/memory under specific load. Harmless for 3 months, it was triggered by a load surge from an unrelated Teams failover plus an unrelated 14% loss of server capacity; 80% of servers went unresponsive, fault-domain isolation failed, and cascading failures caused the global outage.
- Research angle: the team says this is currently very hard for frontier models; SREGym tests whether AI can prevent such outages, with more coming after Thanksgiving.
More from Infra
- Google ships a free fully-local AI notetaker for Mac powered by EmbeddingGemma 2 — dr_cintas · 2026-11-19
- 4x 256GB M5 Ultra Mac cluster with exo serves 32 users at 2000 tok/s on Qwen 27B — StefanoGogioso · 2026-11-19
- Is 262K context worth +0.0011 KLD? A local-deployment KV cache tradeoff question — Distinct-Pie2389 · 2026-11-19
- Chaining DGX Spark boxes with QSFP cables for dumb data-parallel training actually works — austinvhuang · 2026-11-19
- Anthropic's compute spend grows 4x/year, $10B in Q3 alone, set to overtake OpenAI by May 2027 — scaling01 · 2026-11-19
- Thailand data center association cites $18B in GPU/hardware cancellations after SE Asia lending halt — SumitGup · 2026-11-19