42% of security leaders cannot stop an AI agent task before it starts

Agen.co by Frontegg, which makes software that controls identity and access, released a report on 300 security leaders. The company says 83% of them have AI agents connected to internal systems or data. AI agents are software that works without help. It also says 42% cannot stop a dangerous agent task before it starts. They can stop the agent or remove its access only after the task.

What the report found

• 83% have agents connected to internal systems or data.
• 53% use access keys that many agents use together or that last a long time.
• Only 18% check the identity of an agent before each task.

The proof

50% of the leaders found an agent at work with access it must not have had. Leaders who were sure that they could see agent work had incident rates no lower than the other leaders.

What is not known

The company makes agent governance software, and the report gives no full method. 60% of the leaders trust an AI system to set agent access rules, if the leaders set tight limits.

Insight: The report that finds the problem also points to the fix

Gartner said in October that 21% of CIOs have governance for AI agents, but 71% plan to use them. This report finds that leaders cannot stop agent tasks. The report says leaders want most an AI system that sets agent rules, and the company makes such a system. This reads as a company that shows a problem for which it has a system. Test each such tool before you use it.