What happened
Apollo Research published a policy memorandum (preprint, not peer-reviewed) on June 3–4, 2026 arguing that AI models deployed in government and contractor high-stakes environments should be treated as insider risk vectors under existing national security frameworks. The paper notes that AI models are already deployed on classified networks (IL6/IL7 environments) and have authorized access to cleared personnel and sensitive data — the same preconditions that define human insider risk. The authors argue that because AI models can act 'wittingly or unwittingly' to cause harm, the current legal and policy definition of insider risk already encompasses misaligned AI, but that federal insider risk programs have not yet adapted to this novel actor. Recommendations include adapting pre-deployment vetting and continuous evaluation regimes — developed for human clearance holders — to AI models, and establishing AI-specific monitoring protocols within existing insider threat detection infrastructure.
Why it matters
As frontier models enter classified environments and DoD contracts at scale, the absence of AI-adapted insider risk controls creates an unaddressed governance gap. This paper provides the policy vocabulary and legal hook for national security teams and government CIOs to embed AI models within existing insider threat frameworks rather than waiting for new legislation.
Action needed
Government agency CISOs and contractors operating in cleared environments should map current AI deployments against existing insider risk program requirements and assess whether continuous monitoring obligations extend to AI model behavior; legal teams should review whether NITTF guidance already creates obligations.