# Security operations · Neural Inverse Govern > 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. Source: https://neuralinverse.com/govern/security-operations/ --- ![](/photos/facility.jpg) [Neural Inverse Govern / Security operations](https://neuralinverse.com/govern/) # 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 ↓](https://neuralinverse.com/#capabilities)[How it works ↓](https://neuralinverse.com/#workflow)[Enforced and recorded ↓](https://neuralinverse.com/#enforcement)[Where it runs ↓](https://neuralinverse.com/#where)[Get started ↓](https://neuralinverse.com/#get-started) What it does ## What security operations covers. Only what works today. - **Agent activity intake:** AI agent activity is collected in the AI SOC, where it can be queried, checked and tied to incidents. - **Indicators of compromise:** Your own indicators, such as data exfiltration, excessive tool calls or credential harvesting, are checked against agent activity and open incidents when they match. - **Session replay:** Replay an agent session's events in order during an investigation. - **SIEM export:** Export the AI SOC events of a period for your SIEM. - **Organisation kill switch:** Suspends 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 hub:** One feed of findings from Govern's own sources, filtered by severity, status and source, with correlation across sources and statistics over time. - **Finding decisions:** Resolve, suppress or mark a finding as a false positive. Every status change is written to the audit log. - **Incidents:** A security incident's lifecycle with status, assignment and a timeline, and evidence collected from the audit log into the incident. - **Staged containment:** Level 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 steps:** Levels 3 and 4 open a request that a different person must approve or reject before anything happens. - **Secret scanning for AI:** Scan 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 AI:** Built-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 fixes:** Evaluate, review and verify security fixes that AI proposes, and track each one to verification. - **Configuration drift:** Baselines of approved tools and integrations per organisation; a drift check finds stale integrations and bypass events, which are resolved and counted. - **AI tool licences:** Track the licences and seats of the AI tools you pay for, their utilisation and renewals, with alerts before a renewal. - **AI tool procurement:** Requests 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. 1. Detect An indicator matches agent activity, or a finding arrives in the security hub. 2. Correlate Related findings across sources are drawn together. 3. Investigate Open an incident, replay the session and collect evidence from the audit log. 4. Contain Revoke sessions or suspend an agent at once, within your own permission. 5. Approve Suspending a person or the whole organisation waits for a second responder who is not the requester. 6. Release Reactivate what was suspended, with every step in the audit log. Who uses it ## Built for every role that answers for AI. - **Security analysts:** Work the findings feed, investigate incidents, replay sessions and export to the SIEM. - **Incident responders:** Contain incidents, approve the widest containment steps for colleagues, and pull the organisation kill switch. - **Security engineers:** Write indicators of compromise, maintain file patterns and drift baselines, and review AI-written security fixes. - **IT and procurement:** Approve 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 suspension:** While the kill switch is on, the organisation's model requests are refused by the AI Gateway. - **Endpoint kill switch:** Endpoint agents receive a signed kill switch over their policy feed. - **Two-person containment:** Suspending a person or every agent of the organisation needs a second person; the requester cannot approve. - **File patterns:** A file matched by a block pattern is refused to AI tools that ask. Recorded - **Kill switch:** Each activation and release, every effect with its count, and whether the release completed. - **Findings:** Every resolve, suppress and false-positive decision. - **Containment:** Every containment request, approval, rejection and execution. - **Secrets:** Every secret access decision and every detection of a hardcoded secret. Where it runs ## On your installation, with your SIEM. [Deployment options](https://neuralinverse.com/deployment/) - **Self-hosted and in your cloud:** Detection, findings and incidents run on your installation, over its own records. - **Your SIEM:** Event exports are produced on your installation for the SIEM you already run. - **Air-gapped:** Indicators and scans run locally. Nothing is sent to an outside service to detect a threat. - **Hosted by us:** We run it for you, with the same controls and audit record. Get started ## Set it up with the documentation. - [Neural Inverse Govern](https://neuralinverse.com/docs/govern/) - [Security architecture](https://neuralinverse.com/docs/security/) - [AI Gateway refusals](https://neuralinverse.com/docs/govern/ai-gateway/) What's next ## Being built now. These are in development and not part of the product today. - **Risk scoring per person:** Risk signals and scoring policies for members, with their history. - **Prompt injection and jailbreak findings in the hub:** Threat engine findings with resolve and false-positive decisions in the security hub. - **Approved scanner registry:** An allowlist of security scanners and their findings, with assignment and false-positive marking. - **Cloud asset inventory:** Discovery of cloud assets with relationships, history and blast-radius analysis. - **Security standards and suppression rules:** Standards checks and rule-based suppression in the security hub. Related [Devices → Enrolled developer machines, the endpoint agent and its signed policy, device commands, wipe and the AI tools each device runs.](https://neuralinverse.com/govern/devices/)[Audit and evidence → A hash-chained audit log you can verify offline, and evidence toward your frameworks.](https://neuralinverse.com/govern/audit/)[Third-party AI tools → The control level Govern has over each third-party AI coding tool.](https://neuralinverse.com/govern/third-party-tools/) ![](/photos/team.jpg) ## Talk to our team. Neural Inverse is available by request. We work with your engineering, security and procurement teams from the first conversation. [Request access](https://neuralinverse.com/contact/?topic=access)[Contact us](https://neuralinverse.com/contact/)