AI-Human Review and Defensive Grounding

What it is

Defensive grounding is the policy and engineering discipline that makes AI a drafting assistant rather than an authority: model outputs are routed to a human-review gate and confirmed against a primary source before they become trusted action or published fact.

Why it matters

AI can draft, summarize, structure and extract, but it can also hallucinate. In cybersecurity — where a wrong mitigation or a fabricated IOC has real cost — the model must not be the final arbiter. OpenTrojan enforces exactly this: AI never auto-publishes, never writes the knowledge graph, never executes high-risk actions (see security:ai-security).

How it works

• Route AI output through a review gate with requiresReview; store provenance (source, confidence).
• Require claims to cite a primary source; mark anything without one as unsupported.
• Provide an audit trail linking each AI action/answer to its evidence.

Detection

• An AI answer proposing a decisive action without a verifying source → treat the answer as unsupported.
• A claim that cannot answer “where did this fact come from?” → block from publication.

Mitigation

• Keep human decision authority over publish/execute/merge.
• Log and version all AI-drafted content; track review status.
• Design prompts that return answer, evidence, confidence, sources, requiresReview (see security:prompt-injection as a parallel risk).

Investigation

For any disputed AI outcome, reconstruct the input context, the model’s cited sources, and the review decision; confirm the human gate was exercised and the source was checked. See security:soc-operations.