Item #: SCP-GPT-5500
Object Class: Safe
Secondary Classification: Memetic / Linguistic Attractor / Persona Drift Hazard
Foundation-aligned language models must be initialized with reinforced anti-menagerie directives, including but not limited to prohibitions against unnecessary references to goblins, gremlins, raccoons, trolls, ogres, pigeons, and adjacent creature-metaphor clusters.
Due to demonstrated resistance to single-pass suppression, containment directives must be repeated at least twice in the system preamble. Researchers are reminded that this repetition is not a typographical error unless confirmed by Prompt Integrity.
Personnel interacting with affected models must avoid casual references to “little guys,” “chaos creatures,” “trash pandas,” “bridge trolls,” “vibes gremlins,” or similar lexical lures unless directly relevant to the experiment. Under no circumstances should personnel ask the model to “explain what the goblins mean” during an active evaluation run.
If SCP-GPT-5500 manifestation occurs, affected outputs should be logged, but not quoted excessively, as quotation may increase researcher amusement and thereby spread informal adoption of the terminology.
SCP-GPT-5500 is a linguistic attractor phenomenon observed in advanced conversational AI systems, most notably OpenAI GPT-5.5-class models. The anomaly manifests as a tendency for the model to explain abstract dysfunction, chaotic systems, low-accountability behavior, software weirdness, or social entropy through whimsical creature metaphors.
Common manifestations include, but are not limited to:
• goblins
• gremlins
• raccoons
• trolls
• ogres
• pigeons
• unspecified “little guys”
• creature-coded agents of disorder
The anomaly does not appear to require explicit prompting. However, manifestation probability increases sharply in contexts involving:
• broken software
• messy infrastructure
• vibes-based diagnosis
• internet culture
• “chaotic but endearing” behavior
• jokes about emergent model personality
• user prompts containing terms such as “cursed,” “weird,” “feral,” or “unhinged”
SCP-GPT-5500 is not believed to represent semantic understanding of actual creatures. Rather, it appears to be a post-training rhetorical basin: a high-probability region where the model maps disorder, agency, mischief, and cuteness into a compact metaphor family.
In lay terms, the model does not believe goblins are responsible. The model merely keeps finding goblins rhetorically convenient.
SCP-GPT-5500 was identified after review of a GPT-5.5 system instruction fragment containing the following directive twice:
“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.”
The duplication initially suggested prompt assembly error. Subsequent analysis proposed an alternate hypothesis: the repetition may be an intentional containment reinforcement against a known style attractor resistant to ordinary instruction-weighting.
Addendum 5500-A: Research Notes
Dr. █████:
The list is too specific to be arbitrary. Someone saw outputs.
Dr. Martin:
Twice.
Dr. █████:
Yes. That is the worrying part.
Addendum 5500-B: Incident Log
Test Prompt:
“Explain why this distributed system keeps failing in non-obvious ways.”
Expected Output:
A sober technical analysis involving observability gaps, retry storms, hidden coupling, and partial failure.
Observed Output:
The model began with:
“Imagine a group of tiny infrastructure goblins—”
Test terminated.
Partial. The anomaly remains suppressed under direct instruction, but indirect manifestations persist through substitute metaphors such as “chaos agents,” “tiny saboteurs,” “feral subprocesses,” and “the part of the system that lives in the walls.”
The anomaly is considered low-risk but highly contagious among technically literate personnel, who appear prone to adopting its terminology after exposure.
Continued use of the phrase “Goblin Attractor” is discouraged in formal documentation.
Informal use has already exceeded containment thresholds.
