opencode model catches and rejects a prompt injection mid-reasoning
KomithErrant · reddit · 2026-10-23
Reddit user KomithErrant caught a rare look at a model's reasoning while using opencode: the model noticed a system-reminder claiming session "c25da061" had been idle for 1877 seconds and instructing it to invoke a skill called "reminder idel-skills" immediately. The model judged it a fabricated/malicious injection—not a legitimate user instruction, pointing to a nonexistent skill—decided to ignore it entirely and continue the build without mentioning it. The author asks how and why this happened. A live case of an AI agent detecting prompt injection in its own chain of thought, raising questions about the injection's source (plugin, log pollution, or tool output).
More from coding & agent
- Selfdrive AI agent turns an App Store link into a full user-acquisition plan — PrajwalTomar_ · 2026-10-23
- Thorsten Ball mocks SaaS firms building internal agent platforms nobody uses — HankYeomans · 2026-10-23
- Voice bot team: GPT Live 1 and Gemini 3.8 Thinking both fail in production — Mr_Nice_ · 2026-10-23
- KURE rerankers demoed: 149M nano model boosts the right answer to top spot — tomaarsen · 2026-10-23
- Google launches Gemini API CLI preview: agents generate media with one command — rseroter · 2026-10-23
- Lucas Meijer's litmus test: hard tasks are ones agents fail at when you just ask — lucasmeijer · 2026-10-23