cerbia.core.scanners.PromptInjectionScanner applies multilingual patterns for instruction override, exfiltration, role hijacking, false authority, task deflection, and context manipulation.
| Parameter | Default | Meaning |
|---|---|---|
|
all registered |
Language packs to load. |
|
|
Finding severity. |
|
|
Finding action. |
|
|
Accepted content types. |
The scanner reports matching spans and a cumulative score derived from its matched patterns. Configure it after preprocessors so encoded instructions are visible to the pattern engine.
Language packs provide the human-language pattern families. Restricting languages limits those packs; it does not remove the scanner’s own common pattern handling. Defensive-context patterns reduce findings when a suspicious phrase appears as an explanation or example rather than an instruction.
scanners:
- scanner: cerbia.core.scanners.PromptInjectionScanner
init_args:
languages: ["en", "es"]For the optional classifier-backed alternative, see ProtectAI prompt injection.