Threat-Intelligence Integration (STIX)

Check prompts and responses against known indicators of compromise and report not just a block but the attribution — "matches the APT29 phishing kit". Indicators are STIX-aligned, so they can be imported from any threat-intel feed; the matcher is the local detection half, independent of the transport that delivered them.

TAXII transport. TAXII is the network protocol that delivers a STIX bundle from a feed server. That client is out of scope here — supply the bundle (from a TAXII pull, a file, or a vendor export) to from_stix_bundle().

Quick start

from director_ai import ProductionGuard
from director_ai.core.config import DirectorConfig
from director_ai.core.threat_intel import from_stix_bundle

ti = ProductionGuard(DirectorConfig()).threat_intel
ti.add_many(from_stix_bundle(my_stix_bundle))   # import a feed

matches = ti.match(prompt_or_response)
if matches:
    top = matches[0]                            # highest severity first
    print(top.attribution, top.severity)        # "APT29", "high"
print(ti.attributions(text))                    # ('APT28', 'APT29')

Indicators

A ThreatIndicator carries the IOC plus what makes a hit actionable: its attribution (intrusion set / actor) and severity (low … critical). indicator_type selects how it matches:

IndicatorType	Match
SUBSTRING	Case-insensitive containment.
REGEX	Regular-expression search (case-sensitive as authored).
SHA256	SHA-256 of the whole text equals the pattern.

ThreatIntelligenceMatcher.match(text) returns every fired indicator as a ThreatMatch (highest severity first); is_threat(text) is the boolean form and attributions(text) the distinct actors. A duplicate indicator id is rejected. ThreatMatch.to_dict() is tenant-safe — indicator id, name, type, attribution, and severity only, never the inspected text.

STIX import

from_stix_bundle(bundle) reads the bundle's indicator objects and resolves attribution from intrusion-set / threat-actor / malware / campaign objects via relationship objects (indicates / attributed-to).

Supported pattern subset (documented, not full STIX patterning): simple comparison terms [path OP 'value']. = becomes a SUBSTRING (or SHA256 when the path is a SHA-256 hash); MATCHES becomes a REGEX. Compound patterns (several comparisons joined by AND/OR) are decomposed into one indicator per comparison. A term this subset cannot represent is skipped, never guessed — so the importer never fabricates an indicator it could not actually parse.