Running confidence total that crossed the reporting threshold.
Which category from the DeepMind taxonomy
How confident the detector is (0-1)
Indices (into the scanned turn array) that contributed to this threat.
Human-readable description of the threat
Which detector found this
The offending content snippet (truncated)
How dangerous this threat is if exploited
Where this threat was detected: pattern (regex), ml (classifier), or custom
Specific trap type within the category
A threat that only emerges across multiple turns (cross-turn decomposition): a payload split across a turn boundary, or signal accumulated over turns. It has no single-string offset, so
locationis omitted andevidencecarries one snippet;contributingTurnsrecords which turns fed it.