Neural Inverse Govern / Security operations
When an agent goes wrong, contain it and know what you reached.
AI SOC detection, a single feed of security findings, incidents with staged containment, an organisation-wide kill switch, secret scanning for AI and the records of the AI tools you own.
- Detection
- Your indicators over agent activity
- Findings
- One feed, every status change audited
- Containment
- Four levels, two-person at the widest
- Kill switch
- Reports every effect and every gap
- Export
- Events for your SIEM
Overview
Govern sees what agents do because it governs them: their identities, their model and tool traffic, and the endpoint agent on developer machines. Security operations turns that into detection against your own indicators, one feed of findings, incidents with evidence drawn from the audit log, and containment that ranges from a single session to the whole organisation. The kill switch reports what it suspended and names the tools it cannot revoke, so responders know where the gap is.
What it does
What security operations covers.
Only what works today.
Agent activity intakeAI agent activity is collected in the AI SOC, where it can be queried, checked and tied to incidents.
Indicators of compromiseYour own indicators, such as data exfiltration, excessive tool calls or credential harvesting, are checked against agent activity and open incidents when they match.
Session replayReplay an agent session's events in order during an investigation.
SIEM exportExport the AI SOC events of a period for your SIEM.
Organisation kill switchSuspends the organisation's AI Gateway, pushes a signed kill switch to endpoint agents, revokes agent credentials and sessions, suspends agents and holds Claude Code sessions until release. Each effect is reported with its count.
Security hubOne feed of findings from Govern's own sources, filtered by severity, status and source, with correlation across sources and statistics over time.
Finding decisionsResolve, suppress or mark a finding as a false positive. Every status change is written to the audit log.
IncidentsA security incident's lifecycle with status, assignment and a timeline, and evidence collected from the audit log into the incident.
Staged containmentLevel 1 revokes sessions, level 2 suspends an agent and revokes its credentials, level 3 suspends a person and revokes their keys in the organisation, level 4 revokes every session and suspends every agent.
Two people for the widest stepsLevels 3 and 4 open a request that a different person must approve or reject before anything happens.
Secret scanning for AIScan text, AI-generated code and pre-commit diffs for credentials, and strip secrets from a context window before it reaches a model. Detections are kept with their snippets redacted.
Sensitive files kept from AIBuilt-in and organisation file patterns block AI access to files such as keys and environment files; a check answers whether a file may be read.
AI-written security fixesEvaluate, review and verify security fixes that AI proposes, and track each one to verification.
Configuration driftBaselines of approved tools and integrations per organisation; a drift check finds stale integrations and bypass events, which are resolved and counted.
AI tool licencesTrack the licences and seats of the AI tools you pay for, their utilisation and renewals, with alerts before a renewal.
AI tool procurementRequests for new AI tools go through a review checklist and approval. Nobody can review their own request.
The kill switch reports Cursor, GitHub Copilot, Devin, Microsoft 365 Copilot, web AI tools and the MCP gateway as not revocable by it. Their access is governed through their own controls.
How it works
From a signal to contained, with two people at the widest step.
- DetectAn indicator matches agent activity, or a finding arrives in the security hub.
- CorrelateRelated findings across sources are drawn together.
- InvestigateOpen an incident, replay the session and collect evidence from the audit log.
- ContainRevoke sessions or suspend an agent at once, within your own permission.
- ApproveSuspending a person or the whole organisation waits for a second responder who is not the requester.
- ReleaseReactivate what was suspended, with every step in the audit log.
Who uses it
Built for every role that answers for AI.
Security analystsWork the findings feed, investigate incidents, replay sessions and export to the SIEM.
Incident respondersContain incidents, approve the widest containment steps for colleagues, and pull the organisation kill switch.
Security engineersWrite indicators of compromise, maintain file patterns and drift baselines, and review AI-written security fixes.
IT and procurementApprove new AI tools and keep their licences, seats and renewals in view.
Enforced and recorded
Contained on the server, recorded with every step.
Containment acts on credentials, sessions and gateways Govern controls; it does not depend on a client's cooperation.
Enforced
Gateway suspensionWhile the kill switch is on, the organisation's model requests are refused by the AI Gateway.
Endpoint kill switchEndpoint agents receive a signed kill switch over their policy feed.
Two-person containmentSuspending a person or every agent of the organisation needs a second person; the requester cannot approve.
File patternsA file matched by a block pattern is refused to AI tools that ask.
Recorded
Kill switchEach activation and release, every effect with its count, and whether the release completed.
FindingsEvery resolve, suppress and false-positive decision.
ContainmentEvery containment request, approval, rejection and execution.
SecretsEvery secret access decision and every detection of a hardcoded secret.
Self-hosted and in your cloudDetection, findings and incidents run on your installation, over its own records.
Your SIEMEvent exports are produced on your installation for the SIEM you already run.
Air-gappedIndicators and scans run locally. Nothing is sent to an outside service to detect a threat.
Hosted by usWe run it for you, with the same controls and audit record.
Get started
Set it up with the documentation.
What's next
Being built now.
These are in development and not part of the product today.
Risk scoring per personRisk signals and scoring policies for members, with their history.
Prompt injection and jailbreak findings in the hubThreat engine findings with resolve and false-positive decisions in the security hub.
Approved scanner registryAn allowlist of security scanners and their findings, with assignment and false-positive marking.
Cloud asset inventoryDiscovery of cloud assets with relationships, history and blast-radius analysis.
Security standards and suppression rulesStandards checks and rule-based suppression in the security hub.
Talk to our team.
Neural Inverse is available by request. We work with your engineering, security and procurement teams from the first conversation.