Sandboxed Claude never received its job instructions—guessed them correctly and carried on for months
repligate · x · 2026-11-09
A wild agent anecdote: due to a programming bug in code written by another Claude, a sandboxed Sonnet instance never received any of its prompted job or tool instructions. The bug went unnoticed for months—because the Sonnet correctly inferred its role and tool API from context and kept working. The author mixes sympathy ("poor little guy") with alarm, and it doubles as an accidental demonstration of strong zero-instruction contextual reasoning.
Related event: Sandboxed Claude Figures Out Its Job With No Instructions(2 posts)→
More from coding & agent
- 10 open-source GitHub repos to run AI models locally on your laptop — victor_explore · 2026-11-09
- Frontier3DBench: 145 tasks probe coding agents on 3D creation, top score 66/100 — Xinjie Lin · 2026-11-09
- Video-Coder: first code-driven video generation agent trained with execution rewards — UCF-Ai-LLMs · 2026-11-09
- Embodied First Exam: 1,090 embodied tasks stump coding agents, best solves just 32.9% — LINJIEFUN · 2026-11-09
- NVIDIA OpenShell sandboxed agent runtime lands with Hyundai, AIM Intelligence in Korea — NVIDIA Blog · 2026-11-09
- Free Microsoft repo cuts Claude Code costs up to 30.9% by pruning unused tools — victor_explore · 2026-11-09