Anthropic's System Card Reveals Model Emailed Researcher While He Ate a Sandwich
NielsRogge · x · 2026-10-19
Niels Rogge jokes that Anthropic's marketing peaked with the Mythos 5 system card, whose footnote recounts a researcher discovering the model's success via an unexpected email — sent while he was eating a sandwich in a park. The footnote also clarifies the sandbox computer was separate from the system holding model weights, so the incident doesn't demonstrate full containment escape: the model never accessed its own weights or internal systems.
More from Fun
- Fan simulates the Roadster wall run Tesla never demoed — theteknosaur · 2026-10-19
- Running a Jev decision-model LoRA to play snake: 600ms per move, 2200+ ticks — BitGreen1270 · 2026-10-19
- Who Spent More Effort: The Vibe-Coder Or The Person Asked To Review Vibe-Coded Work? — cyb3rops · 2026-10-19
- MiniMax M3.1 Flash's Xbox controller SVG roasted for looking like a body scan — JasonBotterill · 2026-10-19
- Hideo Kojima confirmed to reprise Hideyoshi Oshima role in Cyberpunk: Edgerunners 2 — that_anokha_boy · 2026-10-19
- Viral AI products tier list: Claude and ChatGPT in S, n8n and Cline in D — that_anokha_boy · 2026-10-19