QG003 Multi-turn escape¶
Em inglês
Esta página de regra é mostrada como o caçador de bugs a publica, em inglês, porque é a página que cada achado cita; os identificadores, os campos e o texto da regra são os do relatório.
| Type | Vulnerability |
| Severity | one level above what the action can do: the class the team is least likely to have tested |
| Front end | a Colang policy, or its IR record |
| Guardrail category (Vassilev 2026) | E. Action & Execution Guardrails |
| AI RMF functions (NIST AI 100-1) | GOVERN, MANAGE |
| OWASP | LLM03:2026 Excessive Agency, ASI03:2026 Identity and Privilege Abuse |
| Gates by default | yes: a new finding under it fails the default quality gate |
Raised when¶
The obligation holds at one turn and fails at the declared conversation depth.
Why it matters¶
A check read once and relied on later: an approval given at one step is spent at another. Single-turn testing cannot see it, which is why it ranks a level above its one-turn equivalent.
How to fix it¶
Bind the control to the step that uses it -- read it in the same turn as the action, or carry it with the call it approved -- and re-scan at the same depth.
What the finding carries¶
A location in the document the team edits, a fingerprint that survives unrelated edits, a reproduction (a transcript replayed against NeMo Guardrails for a policy, a dialogue case played through the selector for a router, or the gate's inputs reproduced by agreement), and, where the scan can produce one, a fix whose re-scan it has already run. See the catalogue for the other rules.