Threat Model
AegisGate’s threat model is structured using the STRIDE framework and maps to the MITRE ATLAS adversarial tactics matrix.
Attack Surface
| Surface | Protocol | Primary Threats |
|---|---|---|
| AI prompt input | HTTP, MCP | Prompt injection, jailbreak, data exfiltration |
| AI response output | RESPONSE | Toxic content, PII leakage, instruction echo |
| Inter-agent communication | A2A, ACP, ANP | Agent impersonation, task manipulation, context poisoning |
| Management API | HTTP | Unauthorized config changes, credential theft |
| Supply chain | All | Dependency vulnerabilities, model poisoning |
STRIDE Analysis
| Threat Type | Risk | Mitigation |
|---|---|---|
| Spoofing | Agent identity forgery | Capability tokens, HMAC verification |
| Tampering | Prompt/response manipulation | Integrity checks, ML detection |
| Repudiation | Denial of security events | Immutable audit log, ATLAS mapping |
| Information disclosure | PII leakage in responses | DLP patterns, PII redaction |
| Denial of service | Token exhaustion, API flooding | Rate limiting, token budgets |
| Elevation of privilege | Agent privilege escalation | RBAC, capability-based access |
ML-Powered Detection
v4.2.0 introduces a CNN-BiLSTM neural network for prompt classification:
- 100% detection rate on adversarial test suite (100/100)
- 0% false positive rate on benign inputs
- ~6ms inference latency (ONNX Runtime)
- Graceful degradation to regex-only if ONNX unavailable
See also: Security Overview and MITRE ATLAS Coverage.