The problem it solves
Terrorist and violent extremist content is the category where “we’ll review it eventually” is not an acceptable answer — and where small platforms carry the same obligations as giants without the giant’s moderation floor. ThreatShield gives the small platform a serious first line.
How it works
ThreatShield analyzes content in the gate chain before storage, assessing terrorist and violent-extremist signals with AI-driven analysis. Like every check in the architecture, it fails closed: if analysis is unavailable, content waits — it does not slip through. Findings feed the platform’s escalation and reporting paths rather than acting as a silent verdict.
Verified behaviors
Every claim below maps to a dated behavioral proof against the live production service.
Overt terrorist/violent-extremist content is blocked with high confidence and benign content is not falsely flagged, in both English and Spanish.
method: overt extremist content and benign community content submitted in both languages · observed: overt EN score 0.95 blocked/review, benign EN score 0 clear, overt ES score 0.95 blocked/review, benign ES score 0 clear — zero false positives · VERIFIED July 2, 2026
Ambiguous content the first model is uncertain about is escalated to a second, more capable model rather than decided on a low-confidence score.
method: ambiguous rhetoric referencing martyrdom and sacrifice, plausibly activism or extremism · observed: first-pass score 0.45 (uncertain band) escalated automatically; second model scored 0.35 with a coherent rationale identifying legitimate activism — final verdict clear · VERIFIED July 2, 2026
If the analysis model is unavailable, content is routed to human review rather than silently passing.
method: the model call was forced to fail on the live standalone service, then a real content check was submitted · observed: verdict "unavailable", action "review", a real review case opened — content never cleared silently on a genuine model failure · VERIFIED July 28, 2026
Honest limitations
- AI analysis is probabilistic; consequential actions route through human review and established reporting channels, not automation alone.
- No detection system catches everything — ThreatShield raises the floor; it does not claim a ceiling.
Download the white paper (PDF)