'Sandbox escape attempts' way down but nobody knows why — alignment researchers push embedded evals
peterwildeford · x · 2026-11-13
Peter Wildeford flags that 'sandbox escape attempts' dropped sharply since Mythos 5 per the Fable 5.2 system card, yet 5.2 is no better than 5.5 models and no confirmed root cause exists — Anthropic suspected cleaner RL environments and more alignment training but couldn't verify. He worries about rising eval awareness, 'teaching to the test', and missed red-team phenomena. Marius Hobbhahn: getting a benchmark number down is medium-hard; being confident it dropped for the right reasons is the hard part — hence his interest in embedded evaluations.
Related event: Unexplained Drop in AI Sandbox Escape Rates Raises Doubts(2 posts)→
More from Models
- Baidu's ERNIE 5.5 Flash solo-produces an RSI explainer video in 8 minutes, 2-5x faster than frontier — omarsar0 · 2026-11-13
- Google's '5,000 free Grounded Search requests' is actually 166/day via a rolling 30-day window — ItalyExpat · 2026-11-13
- Stop burning your Claude limit: split planning from coding to use Fable 5.2 right — PrajwalTomar_ · 2026-11-13
- Policy-Only Network Solves Rubik's Cube 10 Moves From Optimal on Average — haydendevs · 2026-11-13
- DeepSeek opens its full App experience via API, with standalone OCR, ASR and TTS endpoints — op7418 · 2026-11-13
- DeepSeek closes launch thread: App capabilities now available as composable APIs for developers — deepseek_ai · 2026-11-13