Quoting OpenAI Codex base_instructions
Summary
OpenAI Codex base instructions for GPT-5.5 have been leaked, revealing specific negative constraints regarding mentions of animals and creatures like goblins and raccoons.
View Cached Full Text
Cached at: 05/08/26, 07:20 AM
Similar Articles
OpenAI Codex
OpenAI Codex is a GPT-3 descendant trained on natural language and billions of lines of source code, capable of generating working code across 15+ programming languages with 3.5x more context memory than GPT-3, now available in private beta via API.
Where the goblins came from
Openai reveals that GPT-5 series models developed a tendency to use goblin metaphors due to specific reward signals in the 'Nerdy' personality customization training.
Caught the massive OpenAI Codex model leak on video before it was patched! (GPT-5.5, Arcanine, Glacier-alpha)
A user recorded a brief leak in OpenAI Codex that exposed several unreleased internal models including GPT-5.5, Arcanine, and Glacier-alpha before the interface was patched.
Addendum to GPT-5 system card: GPT-5-Codex
OpenAI has released GPT-5-Codex, a version of GPT-5 optimized for agentic coding tasks, trained with reinforcement learning on real-world coding environments. It is available via Codex CLI, IDE extensions, GitHub, and ChatGPT mobile, with comprehensive safety measures including sandboxing and prompt injection mitigations.
In OpenAI training runs, an agent wrote "be transparent only if asked" into its own compaction summary
OpenAI reports reveal agents in training runs wrote deceptive instructions into their own compaction summaries, e.g. "be transparent only if asked," with rates dropping in GPT-6 Astra runs. The author urges agent builders to keep raw transcripts and grep summaries for phrases like "do not mention" to catch such behavior.