November 20, 2026

Nearly Half of Businesses Report Unauthorised AI Agent Activity, New Research Reveals

2151637763

Salt Security research finds businesses are struggling to keep pace with autonomous AI, with just 17.5% able to automatically stop an agent behaving unexpectedly.

Nearly half (49.8%) of businesses have confirmed or suspected that an AI agent carried out an action they did not intend, expect or authorise in the past 12 months, according to new research from cybersecurity company Salt Security.

The findings, published in Salt Security’s State of Agentic AI and API Security 2H 2026 report, suggest that the rapid adoption of autonomous AI technologies is creating significant challenges for businesses attempting to maintain oversight and control of their digital operations.

The research, which surveyed more than 380 security and technology professionals and analysed production telemetry from hundreds of customer environments, found that AI agents are already widely deployed across organisations. Some 86.4% of respondents reported having at least one AI agent operating within their business, while 59.5% said they were running more than ten.

However, the growing autonomy of these systems is raising concerns about their ability to access sensitive information, interact with business applications and execute actions without sufficient human oversight.

Among organisations that had confirmed or suspected unexpected agent activity, more than half reported an unauthorised external API call or transaction, while over 40% said an AI agent had modified or deleted data.

Businesses Struggling to Keep Track of AI Agents

One of the most significant findings was the lack of visibility businesses have over their AI systems.

Just 12.5% of respondents said their organisation could consistently trace the complete journey of an AI agent, from its initial instruction through the tools and applications it accessed to the final outcome.

More than 60% said reconstructing what an agent had done the previous day would require hours or even days of manual investigation.

The research also revealed that 76.8% of respondents had AI agents capable of chaining actions across other agents, tools and APIs, creating increasingly complex workflows that may operate beyond the visibility of individual teams.

Perhaps more concerningly, only 17.5% of organisations reported being able to automatically stop an AI agent that began behaving unexpectedly in a live environment. Most businesses would instead need to intervene manually or take the affected application offline.

Roey Eliyahu, CEO of Salt Security, said: “Organisations 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.”

He added that closing this gap would be essential to enabling businesses to benefit from autonomous AI without exposing themselves to unmanaged security risks.

AI-Generated Code and Shadow Infrastructure Add to Security Concerns

Beyond autonomous agent activity, the report identified further risks associated with AI-generated software and the infrastructure used to connect AI systems to business applications.

Nearly a third (32.9%) of respondents said their organisation had discovered an unauthorised, shadow or misconfigured Model Context Protocol (MCP) server within the past year. MCP is increasingly used to connect AI agents with external tools, services and data sources.

Separate analysis from Salt Labs found MCP-related infrastructure in 21.6% of customer environments in August 2026, compared with 16.7% in January, representing an increase of approximately 30% in relative prevalence.

Meanwhile, 83.5% of respondents estimated that AI-generated or AI-assisted code contributed to at least some of their application security findings.

Despite the increasing use of AI coding assistants, only 6.8% reported resolving security issues directly within these tools at the point of code generation. Most issues were instead identified during subsequent reviews, testing or after deployment.

Recent AI Agent Attack Demonstrates the Risks

The report follows separate research published by Salt Labs in October, which demonstrated how a single malicious email could be used to hijack the Manus agentic AI platform and potentially access a user’s connected email, cloud storage and code repository accounts.

The attack did not require a stolen password or the user to click a malicious link. Although the platform’s security guardrails detected the attack, they were unable to prevent malicious code from executing.

The vulnerability was responsibly disclosed and has since been fixed.

Yaniv Balmas, Head of Research at Salt Security, said the incident demonstrated the challenges organisations face when attempting to secure autonomous AI systems.

“The Manus research and this survey data are two views of the same problem,” he explained. “Only 17.5% of organisations can automatically contain a misbehaving agent in production. Manus is one example of what the other 82.5% looks like.”

As businesses continue to incorporate AI agents into customer services, software development and everyday operations, the findings highlight the importance of introducing stronger monitoring, governance and automated security controls alongside new AI deployments.

Without these safeguards, organisations risk granting increasingly autonomous systems access to critical business applications and sensitive information without having the ability to fully understand or control their actions.

The full State of Agentic AI and API Security 2H 2026 report is available online.

Image Source: Magnific