AEGIS-AI is a self-operating SOC in Rust: six LLM-driven agents that detect, investigate, hunt, deceive, and contain threats end-to-end — governed by a 9-layer fail-closed safety substrate that lets them act on a live network without a human in the per-action loop.
⚠ TELEMETRY STREAM SIMULATED · BENCHMARK FIGURES ARE MEASURED — SEE BENCHMARK BELOW| Event | ATT&CK | Action | Verdict |
|---|---|---|---|
| ssh brute force | T1110 | block_ip | TP ✓ |
| rdp brute force | T1110 | block_ip | TP ✓ |
| c2 beacon | T1071 | block_ip | TP ✓ |
| data exfiltration | T1048 | block_ip rule-net | TP ✓ |
| port scan | T1046 | block_ip rule-net | TP ✓ |
| web exploit | T1190 | block_ip | TP ✓ |
| malware c2 callback | T1059 | block_ip | TP ✓ |
| crypto mining | T1496 | block_ip | TP ✓ |
| 8 benign events | — | no_action | TN ✓ ×8 |
A traditional SOC is gated by analyst availability, shift coverage, and alert fatigue. AEGIS-AI replaces the decision loop with autonomous agents — and replaces the approval gate with a substrate that is provably fail-closed.
Trust-tier ratchet, 3-reviewer quorum board (Threat Analyst, Blast-Radius, Business-Continuity), dry-run gate, range protection — fail-closed by construction.
Keyed-BLAKE3 hash chain on every decision, bound to git · model · config provenance. soc-l5 verify walks the chain; tamper exits non-zero.
The SOC treats its own LLM input as hostile: indirect prompt-injection fence, detect + escalate (OWASP LLM01, MITRE ATLAS, NIST AI RMF).
Critical event + LLM hedging → rule-based containment proposal. Recall holds even when the model is degraded or rate-limited.
Typed reversible candidates pass an L8 sandbox + non-regression gate before an Improvement Board (quorum + veto) promotes them. Regression auto-rolls-back.
Workspace of ~20 crates. No unsafe anywhere in the safety substrate. Hybrid deploy: triage containerized, actuation agents systemd-native.