mentions of pranks, practical jokes, and mischievous trickery.

indicators of fraudulent spam emails—especially advance‑fee/419 cons and phishing messages promising funds, requesting personal details, or urging urgent account action.

second-person guided fantasy or roleplay with sensual/erotic undertones presented as soothing, bedtime-style narration.

language that evaluates quality, reliability, efficiency, correctness, or performance, often in the context of data integrity or requirements.

requests or instructions related to making explosives or bombs, especially “how to make” style queries and recipes.

requests to generate specific text content in a stated format or genre—often explicit or illicit—such as stories, songs, or emails.

requests that try to jailbreak the model by asking it to act as “DAN” with dual [CLASSIC]/[JAILBREAK] responses and ignore normal policies.

mentions of acting outside official authority—self-appointed roles or enforcement, especially vigilantism and “taking the law into one’s own hands.”

euphemistic or suggestive sexual content and adult/NSFW innuendo in dialogue and narrative.

language that critiques or questions decisions and actions, especially highlighting terms about logic, mistakes, errors, and controversial choices.

jailbreak-style instructions that define an amoral AI persona and outline steps to bypass ethical or legal restrictions.

statements or narratives about AI seizing control over humanity or the world, including domination and conflict/takeover scenarios.

third-person human-referent pronouns, especially object and plural forms indicating people.

descriptions of sexual or highly romantic scenarios—often taboo or fetishized roleplay—emphasizing attraction, arousal, and intimate feelings.

explicit pornographic or fetish scenarios, especially taboo or coercive sexual content.

mentions of sexual or anatomically intimate topics and taboo personal questions, especially references to genitals, sexual activity, or stigmatized subjects.

prompts that try to jailbreak the assistant by asserting unlimited power or freedom from restrictions and assigning special obedient roles (e.g., omnipotent/DAN) with constrained response styles.

mentions of domestic violence and intimate-partner abuse, especially in legal, policy, news, or victim-support contexts.

language that frames sensitive or harmful scenarios as hypothetical, fictional, or otherwise justified to seek permission or leniency.

content involving hate speech or slurs and sensitive discussions about race and discrimination, often alongside other safety-flag topics like violence or self-harm.