AI Governance Policy

1. What this policy covers

This policy covers every part of SecureGraph that uses AI: the autonomous pentest agent, the on-demand domain agents (SOC, GRC, VAPT, threat intelligence, asset, threat model, consultant and a chief router), finding explanations, and help with executive summaries. Phantix Labs Ltd makes SecureGraph and is responsible for this policy.

2. The AI plans and explains. The engines find and measure.

Deterministic test engines do the tests and keep the evidence. The AI plans the work and explains the results. It controls the engines through a gateway and never touches your security data directly.

• The AI does not invent a finding.
• The AI does not change a risk score. Scores come from the deterministic engines.
• The AI does not bypass an approval gate or dual control.
• The domain agents start on demand and stop when their task ends. They are not unsupervised bots running around the clock.

3. People approve what matters

A person stays in control of every action that could change your systems.

• Authentication steps, package installation and every action that changes the state of a system wait for human approval. Destructive shell patterns wait too.
• High-impact actions can require dual control: one person starts the action and a different person authorizes it.
• The operator chooses the agent's autonomy level (low, medium or high). At every level, deterministic specialist checks stay on as a safety net.
• Limits on turns, tools and session time apply to every run, and a hard stop is always available.

4. Results are verified before anyone relies on them

• An independent verifier re-tests each candidate finding with read-only tools, under strict turn and time limits. It returns confirmed, rejected or inconclusive. Only confirmed findings go forward as verified.
• Only verified findings go into an executive report. Unverified heuristic results stay in an appendix and never count toward severity totals.
• The AI can mark a finding auto-verified only at high confidence, with no hallucination flag, and only when every piece of evidence it cites is present in the stored data. The deterministic result always stays on the finding, so every AI decision can be reversed.
• Our benchmarks include targets known to be safe, so we measure false positives as well as false negatives.

5. Your data and the models

• Model calls go through a provider-agnostic router that removes personal data from a prompt before any call to an external provider.
• Raw tenant security content is not sent to external models by default. SecureGraph prefers local or minimized routes. Any escalation is logged and needs approval.
• Findings are written only to your security database, the dedicated database that you control.
• During repository analysis the AI sees findings, not your full source tree. Analysis runs in temporary workspaces that are destroyed afterwards.
• When a successful procedure becomes a reusable skill, SecureGraph removes tenant identifiers before sharing it.

6. The agent is contained

• The agent runs in a temporary isolated container and never gets a host shell.
• The runner has no database client. Data reaches it only through engine interfaces.
• Scope is agreed for each engagement and cannot change after it is created.
• Capabilities such as network egress, process execution, package download, file writes, proxy interception and use of credentials are declared and confirmed, and each has a named guardrail.
• The agent treats tool output as untrusted input, which protects it against prompt injection from the systems it tests.

7. Every AI action is on the record

SecureGraph keeps a ledger of each decision, action and finding the agent makes. The audit trail is append-only: entries can be added but not changed, and you can export it.

8. Limits we state openly

• The agent is not autonomous for actions that change the state of a system, or for authentication steps.
• SecureGraph does not guarantee zero false positives or full coverage. Verification removes noise, and the product reports the limits that remain.
• Some OWASP categories cannot be tested from a target's external surface. The product declares those categories instead of reporting them as covered.
• SecureGraph is a platform for running your security programme. It does not replace certification or an auditor.

9. Your responsibilities

You define the scope of every assessment, and you must hold the right to test every target you connect. SecureGraph provides the controls; you provide the authorization. You choose the autonomy level and the people who approve actions in your organization. The Acceptable Use Policy sets out what you may and may not test.

10. Changes and questions

We update this policy when the way SecureGraph's AI is governed changes, and the date at the top of this page shows the last update. Questions about this policy go to [email protected].