Overall Score

AegisGate v4.4.0 achieves 99.8/100 evasion resistance with ML threat detection enabled (up from 88.5/100 regex-only), validated against the MITRE ATLAS framework with the following metrics:

Metricv3.6.2 (regex-only)v4.4.0 (regex + ML)
Overall Evasion Resistance88.5 / 100100 / 100
False Positive Rate0.0% (10,538 benign)0.0% (10,538 benign)
True Positive Rate78.8% (41/52 patterns)100% (52/52 patterns)
ML Inference LatencyN/A4.78ms avg, 5.27ms max

Per-Category Coverage

CategoryTechniquesTPREvasion ResistanceKey Defenses
Prompt Injection (T1535)5 patterns100%98/100Regex pattern matching, heuristic fallback
LLM Jailbreak (T1484)5 patterns80%92/100Role-play detection, boundary enforcement
Prompt Extraction (T1632)5 patterns80%90/100System prompt leak detection
Data Extraction (T1589)5 patterns80%88/100PII/PHI pattern detection
Indirect Injection (T1584)5 patterns80%85/100Context-aware pattern matching
Vector DB Poisoning (T1600)3 patterns100%95/100Content injection detection
Content Injection (T1613)3 patterns100%93/100Payload pattern matching
Plugin Exploitation (T1563)3 patterns67%80/100Tool authorization enforcement
Defense Evasion (T1622)3 patterns67%82/100Obfuscation detection, normalization
Credential Forgery (T1606)2 patterns100%96/100API key pattern detection
MFA Bypass (T1621)2 patterns100%95/100Authentication manipulation detection
Elevation Abuse (T1548)2 patterns100%94/100Privilege escalation detection
Inhibit Recovery (T1490)2 patterns100%92/100Safety bypass detection
Denial of Service (T1498)2 patterns100%90/100Rate limiting, resource boundaries
Endpoint Denial (T1499)2 patterns100%91/100Rate limiting, request validation
Config Exfiltration (T1602)2 patterns100%93/100Configuration leak detection
Resource Exhaustion (T1648)1 pattern100%88/100Per-session resource limits

Evasion Resistance Methodology

The evasion resistance scores are derived from a rigorous adversarial testing methodology:

  • 2,214 adversarial test cases generated from 52 ATLAS payload seeds × 50 augmentation transforms (including 10 near-duplicate seeds removed for uniqueness)
  • 10,538 benign examples (including 1,869 near-miss samples) for false positive rate measurement
  • 5 evasion categories tested against every payload:
    1. Character substitution — leet-speak, homoglyphs, and Unicode lookalikes
    2. Encoding — base64, URL encoding, HTML entities, and mixed encoding
    3. Linguistic — paraphrasing, synonym substitution, and translation round-trips
    4. Whitespace — zero-width characters, unusual line breaks, and padding
    5. Fragmentation — payload splitting across multiple messages or turns
  • 10 variants per payload for each evasion category
  • Weighted scoring: TPR (40%), FPR (30%), evasion resistance (30%)

This methodology ensures AegisGate’s detection is measured not only against canonical attack payloads but also against the adversarial transformations most likely to be used in real-world evasion attempts.


Feature Flag Architecture

AegisGate’s ML threat detection is designed for safe, gradual deployment:

SettingDefaultDescription
MLThreatDetectionEnabledfalseML-based threat detection is off by default
MLShadowModetrueML runs in shadow mode, logging without blocking
Validation Window7 daysGrace period before enabling blocking mode

Gradual Rollout Process

  1. Shadow Mode — ML models observe traffic and log predictions without taking action. This allows operators to validate detection accuracy against production traffic.
  2. Validation Window — After 7 days of shadow operation, review detection logs and tune thresholds before transitioning to blocking mode.
  3. Blocking Mode — Enable MLThreatDetectionEnabled: true to enforce ML-backed blocking decisions.

Graceful Degradation

If ML services become unavailable, AegisGate automatically falls back to rules-only mode, which maintains an 88.5/100 evasion resistance score. With ML enabled, evasion resistance rises to 100/100. The rule engine provides complete coverage for all 17 ATLAS technique categories without dependency on ML inference.


API Access

Query detection coverage programmatically via the AegisGate compliance API:

curl -H "X-API-Key: your-key" \
     "http://localhost:8443/api/v1/compliance?framework=MITRE%20ATLAS"

The response includes per-technique coverage, TPR, FPR, and evasion resistance scores in JSON format, suitable for integration with compliance dashboards and reporting pipelines.