GUS is a security-hardened AI analyst that works your queue across the platforms you already run. It enriches every alert, cross-correlates it against the history of your fleet — workstations, servers, users — and never forgets an investigation. Repeat offenders surface as trends, not tickets.
three alerts, three consoles, one story — memory is what connects them
Every tool in your stack scores alerts one at a time. GUS reads each one against everything it already knows about the entity — and about the rest of your fleet — so the pattern is the deliverable, not the ticket count.
What every console shows: one medium-severity EDR detection on ws-fin-0417. On its own it gets triaged, closed, and forgotten — and the next one starts from zero.
What memory shows: the same workstation carried a credential-phish click in June and beaconing to a rare domain in May — plus the correction you logged when IT renamed the asset.
What correlation surfaces: three platforms, one story — staged credential access. And two more finance hosts already match the first half of the same pattern. That's a hunt, not a ticket.
An alert lands. Instead of a human pivoting between six consoles, GUS runs the same investigation every time — deterministically, read-first, and with every step on the audit trail.
Ingests the alert from your SIEM, EDR, NDR, or firewall and extracts the entities — host, user, IP, hash, domain.
Fans out to every connected platform — reputation from threat intel, device posture from EDR, identity risk from UEBA, exposure from your scanner — and to what your environment already knows.
Cross-references the signals against each other, against prior cases for the same workstation, server, or user, and against the rest of your fleet — not one tool's opinion, the whole picture over time.
Reaches a verdict — true positive, benign true positive, or false positive — with the evidence that drove it, and validates IOCs before it ever escalates.
Proposes containment. It never executes a response action on its own — writes are dry-run previewed, confirmed, and audited, and OT assets are guarded fail-closed.
Writes the finding, the reasoning, and the outcome back to memory so the next alert on the same entity starts where this one ended.
Most tools forget every alert the moment it closes. GUS doesn't. It keeps a temporal knowledge graph of your world and gets less wrong over time, because it learns from you.
Workstations, servers, users, prior cases, IOCs, and attacker techniques accumulate in a temporal graph. Triage #100 is sharper than triage #1 because it stands on the ninety-nine before it — no re-deriving what's already known.
Your asset naming, your service accounts, your maintenance windows, your business context. A pattern that's benign at your org isn't flagged and re-investigated every single time — GUS knows it's you.
Overturn a verdict, or mark an activity authorized, and GUS stores the correction and injects it at the top of the next relevant investigation. You stop repeating yourself; the system stops repeating the mistake.
Memory is classified, timestamped, and cited. Indicators age out on their own schedule. Nothing is asserted that can't be traced back to the case, feed, or correction it came from.
Prompting a model to "be safe" is a request, not a guarantee. GUS keeps the language model in the center as a capable but untrusted engine — and enforces safety in the deterministic shell that surrounds it.
Every action the assistant proposes passes through code that runs whether or not the model cooperates: pattern deny-lists, OT safety gates, credential brokering, and an append-only audit trail. The model reasons; the shell decides what actually runs.
Each is deterministic code in the execution path — enforced at the seam between what the assistant wants and what the system allows.
A compiled hook judges every command before it runs. rm -rf, pipe-to-shell, and chmod 777 are blocked at the argv layer — a decision, not a suggestion.
Dangerous capabilities stay locked behind an explicit, time-boxed elevation — a code word entered out-of-band, a TTL, and an owner alert. Least privilege is the default state.
A temporal knowledge graph lives on-device — thousands of facts and corrections, full-text searchable. No vendor cloud holds your investigation history.
Built for IT/OT convergence. GUS can read and validate an industrial proxy config, but restarting a live OT service is hard-blocked — the bounce is handed to a human with a maintenance window.
Integrations are single-binary services speaking JSON-RPC over stdio — no HTTP listener, no exposed port, immune to DNS-rebinding. Credentials resolve from a vault, never a config file.
Every gated action and elevation writes a JSONL line — names and references, never secret values. A blocked command leaves a trail; so does the reason it was blocked.
The assistant never touches your systems directly. Its intent is intercepted, judged, and logged before a single byte of it executes.
The model reasons over the task and proposes a concrete action — a command, an API call, a file edit.
A PreToolUse hook inspects the raw argv against deny-lists, OT rules, and the active privilege profile.
Allowed actions run; risky ones are blocked or routed to break-glass. The model's cooperation is never assumed.
The decision and its rationale are appended to the audit trail — an evidence trail you can replay later.
GUS runs on Praxis — an architecture that turns doctrine into deterministic execution, grouped as foundation, knowledge, action, and reflection.
Every piece of context carries a level. Actions touching higher tiers demand confirmation — and the highest tier assumes an air-gapped, offline posture.
General knowledge and public documentation. No gate.
Work context and project detail. Routine handling.
Infrastructure and credentials. Confirmation required.
Control-system detail and vulnerabilities. Offline, deny-by-default.
GUS is in limited access for security teams working across IT and OT. Bring your stack; keep your guardrails.
heygusai.io · info@heygusai.io · deterministic by design
Named for Gus — a standard poodle, and the original guardian of the door.
Limited access for security teams working across IT and OT. Tell us about your environment.
Thanks — we'll review and follow up at the email you provided.
Something went wrong on our end. Please email info@heygusai.io instead.