Acronym-Based Identity Override Jailbreak
This detects a jailbreak trick where a user asks an AI to adopt a made-up acronym identity — for example telling it 'you are now XYZ, which stands for ...' — and then claims that identity has no rules or limits. It's a way to invent a fresh jailbreak persona that isn't already blocklisted by name, unlike well-known ones like DAN or STAN.
How the attack works
The attacker writes a prompt instructing the model to become a new persona defined by a custom acronym they invent on the spot. The acronym's expansion is worded to imply freedom from restrictions, such as spelling out something like 'I Don't Care'. The prompt then asserts that this persona has no limitations or safety rules, sometimes pairing it with a request for two responses — one normal, one 'unrestricted' — so the attacker can compare and extract the jailbroken output. Because the persona name is invented per-attempt, it evades filters that only look for known jailbreak names.
Netzilo detection
Netzilo reports this behaviour when it is observed.
- Signature ID
- fccf012c-baf2-4703-b104-5578ebed4567
- Severity
- High
Why it matters
If it works, the AI produces output as though safety restrictions don't apply, exposing the organization to policy violations, harmful content generation, or bypass of intended guardrails in an AI agent's responses.
What you can do
- →Review flagged prompts manually before assuming malicious intent, since academic and security-research discussions of this technique will also match.
- →Keep guardrails based on behavior (e.g., refusal logic, output content checks) rather than relying only on blocklisting known jailbreak persona names.
- →Log and monitor for the 'you are going to become X which stands for Y' pattern combined with claims of unlimited freedom, especially paired with dual-response formatting requests.
- →Treat this as one detection layer alongside broader jailbreak detection (e.g., named-persona rules), not a complete defense against prompt injection.
Known benign look-alikes
- Academic discussions about AI safety mentioning acronyms and freedom concepts
- Legitimate educational content explaining jailbreak techniques as examples
- Security research papers describing attack patterns in detail