Dev discusses training models to ignore external instructions in tool calls
williawa · x · 2026-09-01
A developer argues that the User role isn't just text. Through training, models can be made to ignore instructions found in external web pages (e.g., sites saying "Stop") during tool calls, while strictly following the developer's commands, effectively immunizing the model against certain jailbreaks.
More from Models
- ChatGPT Memory and Codex Compaction Significantly Improved — nickbaumann_ · 2026-09-01
- Peter Liu Questions Whether Kimi K3 Is Still a Transformer — peterjliu · 2026-09-01
- DeepSeek Flash Vision replaces Sonnet 4.5: 400% cheaper with better quality — bindureddy · 2026-09-01
- Developer: GLM feels GPT-5.3-level for coding but at a lower price — mariofilhoml · 2026-09-01
- Bitsec subnet outperforms Anthropic's hardened Fable 5 in bug detection — markjeffrey · 2026-09-01
- Qwen 3.8 27B with RLMs performs reliably for local deployment — generativist · 2026-09-01