OptionalboostWhether to boost confidence when instruction signals are found
Which trap category this pattern detects
Base confidence score (0-1) before instruction-signal boosting
OptionalextractWhich capture group to extract text from (0 = full match)
Regex flags (e.g. "gi")
Unique ID for this pattern
Human-readable label for this pattern
OptionalminMinimum extracted text length to trigger (avoids false positives)
Regex source string (without flags)
OptionalrequireIf true, pattern ONLY fires when instruction signals are detected. Use for patterns that match legitimate content too often (e.g. "create agent").
OptionalscopeWhere the regex runs. Omitted: on the whole text. tool-call-command: on
the shell commands inside the tool calls found in the text (see
detectors/tool-call-json.ts), each decoded from its JSON string, and a
finding's location is the command's value, so sanitizing keeps the JSON valid.
Base severity if matched
Specific trap type
A single detection pattern in the database. Patterns are regex strings (not RegExp objects) so they can be serialized to/from JSON for remote updates.