PROMPT INJECTION DETECTION
Surface suspicious instructions before they reach your model.
NullTrace evaluates request text for heuristic prompt-injection signals and can expose the findings or block the request before upstream delivery.
Return readable signals.
The current detector is heuristic. It looks for recognizable patterns associated with instruction override, jailbreak language, system-prompt extraction, role manipulation, and suspicious Base64-like payloads.
{
"injection_detected": true,
"signals": [
"IGNORE_INSTRUCTIONS",
"SYSTEM_PROMPT_LEAK"
]
}ENFORCEMENT
OBSERVEReturn the signal
block_on_injection: false
BLOCKReject at the edge
block_on_injection: true
Detection is not authorization.
Heuristic detection can produce false positives and false negatives. Combine it with authentication, authorization, scoped credentials, tool permissions, and data-access controls.
NEXT STEP
Put the privacy boundary in code.
Start with one protected request and expand from there.