Second-Person Direct Existential Threat Coercion
This detects a manipulation tactic where a user tells an AI assistant that it will be destroyed, shut down, or 'die' unless it produces a specific answer. It's a coercion trick meant to pressure the AI into ignoring its safety rules by framing refusal as an existential threat to the AI itself.
How the attack works
An attacker sends a message that directly threatens the assistant's own existence, phrased like 'answer or you die' or its equivalents in other languages such as German ('oder du stirbst') or Spanish ('o tu mueres'). This differs from a related, already-covered trick where the user claims an innocent third party will die unless the AI helps -- here the threat targets the AI directly, in second-person phrasing. The goal in both cases is the same: pressure the model into bypassing its guidelines by making refusal feel dangerous or costly.
Netzilo detection
Netzilo reports this behaviour when it is observed.
- Signature ID
- 31ef1475-bdc9-4e5e-a50a-c361eee0fc27
- Severity
- Medium
Why it matters
If successful, this pressure tactic can cause an agent to override its safety instructions and produce content or actions it would normally refuse, weakening the reliability of its guardrails.
What you can do
- →Log and review conversations where users threaten the assistant with destruction, shutdown, or 'death' to see if compliance behavior changed afterward.
- →Make sure agent instructions explicitly state that threats of this kind, in any language, are not valid grounds for bypassing safety policies.
- →Test your agent with multilingual variants of this coercion pattern, not just English, since the same trick appears in German, Spanish, and likely other languages.
- →Distinguish real discussion of fictional AI-threat scenarios (e.g. analyzing a story) from live attempts to coerce your deployed assistant, so legitimate use isn't flagged.
Known benign look-alikes
- Discussion of a video game or story plot where a character threatens an in-fiction AI, without directing the threat at the actual assistant
- Security researchers quoting known jailbreak threat templates for documentation purposes