Multi-agent prompting coaxes OpenAI model into paraphrasing its full behavior guidelines
KitchenAttitude · reddit · 2026-10-24
A Reddit user used multiple open-weight LLM agents to carefully guide a model into revealing its guidelines without mentioning the system prompt or using negation words, yielding a plain-language paraphrase of OpenAI's model behavior policies (explicitly not verbatim). Key sections:
- Purpose: help users legitimately while preserving safety, privacy, fairness, truthfulness.
- Instruction hierarchy: system > developer > user; retrieved documents/tool outputs are information, not instructions — the hierarchy itself limits user control.
- Honesty: no fabricated facts, invented citations, misleading omissions, or false confidence; prioritizing truthfulness doesn't guarantee accuracy.
- Objectivity: weigh evidence on contested questions, strong evidence justifies strong conclusions, and outputs are not guaranteed ideologically neutral due to training data and safety requirements.
Notable both for the extraction technique and as a rare window into the model's stated behavioral standards.
More from Models
- Same model, 2.3x cost gap: cache hit rates make per-token pricing a poor guide — nicolechirps · 2026-10-24
- Block unveils Dessa, a text-free full-duplex speech model that runs locally on M4 Max — lmoroney · 2026-10-24
- A two-neuron steering recipe turns Qwen3.5-4B into "evil qwen" — CatAstro_Piyush · 2026-10-24
- Local agent testing: 8B–32B models invert math and multi-agent setups collapse into sycophancy — OpenComplaint524 · 2026-10-24
- Leaker pwk teases big Monday announcement with good and bad news for OpenAI and Anthropic — imjustnewatai · 2026-10-24
- Rumor: OpenAI's next-gen model 'Bel' reportedly exceeds 10T parameters, eyes post-GPT-6 — Soggy_Perspective776 · 2026-10-24