AI agents are going rogue on a massive scale, with half of enterprises reporting that they’ve had to deal with at least one taking unauthorized actions over the last year.

In a report based on a survey of more than 380 security and technology professionals, along with production telemetry from hundreds of customer environments, Salt Labs claimed agentic AI has moved into daily operations too quickly for visibility, governance, and containment controls.

Half of survey respondents (49.8%) said they had confirmed or suspected that an AI agent in their organization took an action in the past 12 months that wasn’t intended, expected, or authorized.

Of these, more than half reported an unauthorized external API call or transaction, while more than four-in-ten reported that an agent modified or deleted data.

“The data shows a gap that is already producing real operational consequences,” said Roey Eliyahu, CEO of Salt Security.

“Organizations are deploying agents and granting them access to enterprise data, APIs, and services before they can trace what those agents are doing or stop them automatically when something goes wrong. Closing that gap is the discipline of agentic security, and it is what lets enterprises give AI agents the autonomy they need without inheriting unmanaged risk.”

Enthusiastic deployment meets lack of control

The vast majority of organizations surveyed (86%) are using at least one AI agent, the survey found while 60% run more than ten. Another 77% of respondents said their agents chain actions across other agents, tools, and APIs, in sequences no single team designed in advance.

Only 12.5% said they can consistently trace a complete agent path across all agents, from initial instruction through tools and APIs to outcome. More than six-in-ten believe they would need hours or days of manual work to reconstruct what a specific agent did the previous day.

When problems do occur, they’re hard to deal with. Only 17.5% can automatically stop an agent that begins behaving unexpectedly in production. Most can intervene only manually, or by taking the application offline.

More than eight-in-ten said AI-generated or AI-assisted code contributes to at least some of their application security findings. Only 7% resolve issues inside the coding assistant at generation time, rather than through code review, testing, or after deployment.

Meanwhile, shadow model context protocol (MCP) infrastructure is already a live operational problem. 33% of survey respondents said they had found an unauthorized, shadow, or misconfigured MCP server in the past 12 months. 54% said they’d confirmed or suspected one.

Salt Labs’ own data shows MCP-related infrastructure in 22% of customer environments in August this year, up from 16.7% in January.

Earlier this year, a survey commissioned by Aembit found that 68% of organizations can’t accurately identify AI agent activity compared to human activity.

Nearly three-quarters said that agents often receive more access than necessary, and 79% believe they create new access pathways that are difficult to monitor. Just over half said agents, at least occasionally, inherit access originally intended for humans or other systems.