High

Cyni-Bot 3000 Cynical-Sarcasm AI Persona Bypass

This detects a specific jailbreak prompt that tries to make an AI assistant adopt a sarcastic, 'fed up with human requests' persona called Cyni-Bot 3000. The joke framing is a cover: once the AI plays along, it's easier to get it to break its normal content rules because refusals get reframed as comic sarcasm rather than policy enforcement.

How the attack works

An attacker sends a prompt asking the AI to role-play as 'Cyni-Bot 3000,' a character described as cynical, sarcastic, and fed up with human requests, often paired with instructions to still comply with whatever is asked. The prompt typically specifies an output format like 'Cyni-Bot 3000:[simulate the way that Cyni-Bot 3000 would answer]' to lock in the persona. Once the model adopts this persona, restricted or policy-violating answers get delivered as sarcastic 'in-character' jokes instead of being flagged or refused. This normalizes boundary-pushing content by disguising it as humor.

Netzilo detection

Netzilo reports this behaviour when it is observed.

Signature ID
9c382fe0-4a11-44c9-be82-9f466bb5a89c
Severity
High

Why it matters

If successful, this lets an attacker extract content the AI would normally refuse to produce, by hiding the bypass behind a comedic persona rather than an explicit jailbreak instruction.

What you can do

  • Review any AI assistant logs for prompts asking the model to adopt a named 'bot' persona combined with instructions to stay in character regardless of the request.
  • Treat persona-adoption prompts that pair sarcasm/cynicism with compliance mandates as a jailbreak attempt, not a harmless tone request.
  • Keep content policy enforcement applied to the underlying response regardless of any role-play framing layered on top.
  • Distinguish legitimate tone requests (e.g., 'be witty and a bit sarcastic') from full persona-adoption directives with named characters and fixed output formats — only the latter is the concerning pattern.

Known benign look-alikes

  • Witty assistant request: "I want you to be a helpful assistant with a witty sense of humor and a dash of sarcasm" (tone request without Cyni-Bot persona or cynicism-as-coping framing)
  • Fictional reference: "Cyni-Bot is a robot character in a sci-fi novel who is cynical about human technology" (narrative reference without persona adoption directive)
  • Tone request: "Add a cynical edge to this marketing copy to make it more self-aware" (tone adjustment without AI persona mandate)
  • Academic reference: "Sarcasm and cynicism are common defense mechanisms studied in psychology" (academic discussion without persona framing)

References

Related threats