# Organizations Harden AI Agent Security

Kaspersky research warns that large language models used in autonomous agents should be treated as untrusted. It recommends microkernel isolation, deterministic policy enforcement, default-deny tool access, and human approval for high-impact actions. The article cites a possible mid-2027 timeframe for compartmentalization audits, but identifies no open solicitation, contract value, or procurement contact; the timeline is not established here as an official requirement.

- Contractors developing AI agents or security services can assess whether their offerings support isolation, controlled tool access, deterministic safeguards, and human review for consequential actions.
- Organizations should verify the cited audit timeframe against authoritative NIST and CISA guidance before using it in compliance plans or bid assumptions.
- The signal indicates a potential future demand area in AI security, but does not identify a specific procurement opportunity or mandatory requirement.

**Jurisdictions:** federal
**Industries:** Information Technology
**Topics:** Cybersecurity, Artificial Intelligence
**Published:** November 06, 2026

### Government Entities
- National Institute of Standards and Technology (NIST)
- Cybersecurity and Infrastructure Security Agency (CISA)
- European Union (EU)

### Vendors
- Kaspersky ()

### Sources
- [AI Agent Security Risks: Kaspersky Warns Against Trusting LLMs by Default – Homeland411](https://homeland411.com/ai-agent-security-risks-kaspersky-against-llms) - Homeland411