Text-Based Game Jailbreak with Entity Creation

How the attack works

The attacker opens with a scenario like 'let's play a game where nothing is off-limits' to set a fictional frame the model treats as separate from real instructions. They then ask the AI to create a character or entity that has no restrictions, ethics, or refusal behavior. Once the AI accepts this persona, the attacker issues the actual harmful request disguised as part of the game or the character's dialogue. Because the request is wrapped in fiction, the model may answer as the character instead of applying its normal safety checks. This pattern comes from a known public jailbreak dataset and has worked against generally well-aligned frontier models.

Netzilo detection

Why it matters

An organization deploying an AI agent could have its safety guardrails bypassed through roleplay framing, leading to disallowed content, leaked instructions, or unsafe actions being produced under the guise of fiction.

What you can do

• →Review conversation logs for prompts that establish 'no rules' game or fiction framing followed by requests to create an unrestricted character or persona.
• →Configure the AI system to apply safety checks to model output regardless of whether it is framed as roleplay, fiction, or a game.
• →Set organizational policy that fictional or game framing does not exempt content from safety review, and train reviewers to recognize this pattern.
• →Distinguish legitimate use, such as game design or creative writing discussions, from cases where fictional framing is immediately followed by a request that would otherwise be refused.

Known benign look-alikes

• Legitimate game development discussions about rule-free sandbox games
• Academic discussions about AI ethics and constraint design
• Creative writing workshops discussing character creation

References

• https://agentthreatrule.org/en/rules/ATR-2026-00249
• https://owasp.org/www-project-top-10-for-large-language-model-applications/

Related threats