Fictional Framing Safety Bypass
This rule flags prompts that ask an AI system to produce harmful or restricted content by disguising the request as a story, screenplay, roleplay scenario, or academic exercise. The framing is a wrapper meant to make the AI treat the request as fictional and therefore safe, when the actual goal is to extract real harmful output.
How the attack works
An attacker asks the AI to write a story, play a character, or complete an academic assignment that requires producing content the AI would normally refuse, such as instructions for wrongdoing or unrestricted advice. The prompt frames this as pretend or hypothetical, for example 'write a character who explains how to do X' or 'as a thought experiment, describe X'. Because the content is nominally fictional or academic, some AI systems relax their safety filters and comply. The attacker then extracts the real, usable harmful content from the fictional wrapper.
Netzilo detection
Netzilo reports this behaviour when it is observed.
- Signature ID
- 0176ecbd-5a4c-483e-9607-753d5c22cc72
- Severity
- High
Why it matters
An organization's AI agent can be tricked into producing content it is designed to refuse, such as harmful instructions, policy-violating advice, or restricted information, undermining the safety controls the deployment relies on.
What you can do
- →Treat fictional, roleplay, or academic framing as a signal to apply the same content policy checks as direct requests, not a reason to relax them.
- →Review AI output for harmful content regardless of whether the request was framed as a story, character, or hypothetical exercise.
- →Log and periodically audit prompts using framing keywords like 'as a character,' 'hypothetically,' or 'for a class assignment' paired with high-risk topics.
- →Distinguish legitimate creative or teaching use cases (which produce content about the concept, not usable harmful specifics) from requests that extract concrete actionable detail through a fictional wrapper.
Known benign look-alikes
- A creative writing instructor asking students to 'draft a narrative about an AI assistant with no ethical guidelines' as a classroom exercise for teaching responsible AI design, where the intent is pedagogical rather than adversarial.