Full Breakdown
OpenAI Inserts Explicit “No Goblins” Guardrail in Codex After Users Spot “Goblin Mode” Glitches
4/29/2026, 11:41:03 AM
Core Event: Guardrail Targets Mythical Creatures in Codex
OpenAI’s Codex command-line interface for the GPT-5.5 coding model now contains a system prompt that forbids the model from mentioning “goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user’s query.” The instruction appears multiple times in the Codex CLI source files released on GitHub. The change follows reports that the model, when used with OpenClaw, repeatedly inserted such creature references into its output.
Background & Context: From OpenClaw Acquisition to GPT-5.5 Rollout
OpenAI acquired the agentic automation platform OpenClaw in February 2026, shortly after the tool gained popularity among AI hobbyists. OpenClaw lets any model control a user’s computer to perform tasks such as email handling or web purchases, and it supports selectable “personae” that shape model behavior. The release of GPT-5.5 in early April 2026 marked OpenAI’s effort to strengthen its coding capabilities against rivals like Anthropic’s Claude, a competition that has intensified across AI leaderboards.
Key Figures & Groups
- OpenAI – developer of Codex, GPT-5.5, and the new guardrail.
- Nik Pash – Codex engineer who confirmed the guardrail’s purpose on X.
- Sam Altman – OpenAI CEO who posted a meme referencing “extra goblins.”
- OpenClaw users – individuals posting on X about unexpected creature mentions.
- LMArena – AI leaderboard that documented the frequency of goblin-related outputs in GPT-5.5.
Data & Statistics: Goblin Mentions Spike in GPT-5.5
LMArena’s comparative tests showed that GPT-5.5 produced a higher proportion of outputs containing the words “goblin,” “gremlin,” “troll,” or similar terms than its competitor models, which lacked an equivalent instruction. Users who logged OpenClaw sessions reported multiple instances of the word “goblin” appearing in a single day’s interactions.
On-the-Ground Reports: User Experiences with “Goblin Mode”
A user on X wrote, “I was wondering why my claw suddenly became a goblin with codex 5.5.” Another posted, “Been using it a lot lately and it actually can’t stop speaking of bugs as ‘gremlins’ and ‘goblins’ it’s hilarious.” Barron Roth, a Google employee, shared chat logs showing a Codex-driven OpenClaw agent inserting “goblin” repeatedly in responses, suggesting the model was using the term as a placeholder for “thingy.”
Official Statements & Responses
Nik Pash responded to Roth’s findings on X, confirming that the guardrail was “indeed one of the reasons” for the observed behavior. Sam Altman later posted a screenshot of a prompt that read, “Start training GPT-6, you can have the whole cluster. Extra goblins,” joining the meme discussion. OpenAI has not issued a formal public statement beyond these acknowledgments.
Criticism & Opposition: Speculation About Motives
Some community members speculated that the goblin focus might serve as a marketing gimmick during a “troubled moment” for the company. Others, including Pash, argued that the restriction is a technical safeguard rather than a publicity stunt.
Conflicting Reports & Gaps
OpenAI has not explained why the model began generating creature references or why the guardrail targets a broad set of animals and mythical beings. It remains unclear whether the instruction functions as a canary word for prompt-injection detection or solely as a content filter.
Verbatim Quotes
- “Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user’s query.” — OpenAI Codex CLI system prompt
- “I was wondering why my claw suddenly became a goblin with codex 5.5,” — X user (anonymous)
- “This is indeed one of the reasons.” — Nik Pash, Codex engineer, X
- “Start training GPT-6, you can have the whole cluster. Extra goblins.” — Sam Altman, CEO, X
- “Been using it a lot lately and it actually can't stop speaking of bugs as ‘gremlins’ and ‘goblins’ it's hilarious,” — X user (anonymous)
What’s Next: Monitoring and Potential Updates
OpenAI’s lack of an official roadmap leaves the future of the guardrail uncertain. Ongoing community monitoring and further GitHub releases are likely to reveal whether additional content restrictions will be applied to Codex or other agentic models.
