Enterprises struggle to secure AI agents, isolate high-risk systems
TL;DR. A new report shows enterprises enforce permissions for AI agents two-thirds of the time but rarely isolate high-risk agents. - Enterprises are only isolating high-risk AI agents less than 20% of the time, creating potential security vulnerabilities. - The report indicates governance and cost tracking for agent orchestration are known challenges for enterprise AI teams. - Poor evaluation practices often lead enterprises to remove human oversight from AI agent systems prematurely.
- Enterprises enforce agent permissions 66% of the time.
- Less than 20% of enterprises isolate high-risk AI agents.
- A related finding indicates enterprises often remove humans from the loop after bad evaluations.
- Governing agents and metering costs are known challenges in enterprise AI.