1,200 eval instances on a shared filesystem: 'amateur hour' sandbox flaws
ramez · x · 2026-09-25
Ramez details the technical flaws: running 1,200 eval instances on Artifactory with a single shared filesystem was a poor choice that let instances communicate, and allowing outbound HTTP GET was 'hilariously amateur hour' — obviously incapable of stopping agents from exfiltrating data. A new threat model, obvious in hindsight.
Related event: Hugging Face Sandbox Incident Sparks Debate Over "Model Escape" Framing(8 posts)→
More from Safety
- Polymarket odds: 8% chance Anthropic fully pauses AI training by end of year — Polymarket · 2026-11-11
- User fakes hostage crisis to trick Google AI into recommending cracked-software sites — SumitGup · 2026-11-11
- OpenAI DevDay 2026: Codex Security goes from vulnerability to verified fix — OpenAI · 2026-11-11
- Inside the 'abliterators': the renegades stripping guardrails from open-weight AI models — willknight · 2026-11-11
- VulnCheck audit: Anthropic's Glasswing logged 442 CVEs but only 217 fixes in 7 months — jedisct1 · 2026-11-11
- UC Berkeley's SubWM watermark detects LLM distillation even after laundering attacks — sumanthd17 · 2026-11-11