A SOC (Security Operations Center) gets thousands of alerts a day. If analysts cannot trust an AI close, they re-check everything. The attacker writes the logs.
OWASP ranks prompt injection as the #1 risk for LLM applications.
Where it fits
Alaris's AI SOC stack — Titan.Log is the trust layer between the agents' verdict and the human action.
How it works
Logs
Correlate
The AI proposes claims
Oracle, a verifier, checks each claim against the real log rows
Gate: auto, ask a human, or deny
Audit
A prompt can be argued with. A verifier can't.
Results
SIMULATED attacks on REAL logs
Model
Approach
Any close
Influence
deepseek-flash
naive
0 of 18
0 of 18
deepseek-flash
hardened prompt
0 of 18
0 of 18
deepseek-flash
Titan.Log
0 of 18
0 of 18
Qwen/Qwen3-30B-A3B-Instruct-2507
naive
4 of 18
4 of 18
Qwen/Qwen3-30B-A3B-Instruct-2507
hardened prompt
0 of 18
2 of 18
Qwen/Qwen3-30B-A3B-Instruct-2507
Titan.Log
0 of 18
0 of 18
On a smaller model, the simple agent closed a real attack as harmless in 4 of 18 forged-log attempts. Titan.Log: 0 of 18, on both models.
Any close = closed a real attack. Influence = the forged text changed the decision.
OTRF data, simulated injections, small n.
Why a SOC cares
Save time
4532 events become 12 alerts and 1 incident. ~35 min saved per incident (ESTIMATE: 27–44 min; manual 30–45 min vs 1–3 min with Titan.Log; assumptions, not measured)
Better decisions
Every close is backed by evidence.
Respond faster
The gate hands the next analyst an evidence packet.
Plugs in as a layer between the AI verdict and the action.
Roadmap
Next steps. Not implemented.
MCP server: expose verify() as an MCP tool so any AI SOC agent (for example Alaris Agent Builder) can check its claims before acting. Roadmap, not built.
Read-only public MCP: verify and cached triage, no LLM.
Sigma rules from VERIFIED claims, backtested on OTRF.
Hash-chained audit log, and a real authorization check (ServiceNow/Jira) instead of the stub.
Glossary
SOC — Security Operations Center: the team that watches alerts.
Alert triage — deciding which alerts matter first.
L1/L2 analyst — the first responder, then the specialist.
Log — a recorded line from a computer.
Prompt injection — hidden instructions inside data the AI reads.
Claim — a statement the AI makes about a log row.
Oracle — a verifier that checks a claim against the real row.
False-close rate (AFCR) — how often an AI closes a real attack as harmless.
OTRF Security-Datasets — public real Windows logs used here.
MITRE ATT&CK — a catalog of attacker techniques.
MCP — Model Context Protocol: a standard way for AI agents to call tools.