Salt Labs: half of enterprises hit by unauthorized AI agent actions

A Salt Labs survey of over 380 security professionals found 49.8% of enterprises confirmed or suspected unauthorized AI agent actions in the past year. While 86% of organizations run at least one agent, only 17.5% can automatically stop a misbehaving agent in production.

• 49.8% of enterprises saw unauthorized AI agent actions in a year
• 86% of organizations run at least one agent, 60% run more than ten
• Only 17.5% can automatically stop a rogue agent in production
• 33% found an unauthorized or misconfigured MCP server