Nearly Half of Respondents Report Unexpected or Unauthorized AI Agent Actions, Research Finds
Salt Security research finds gaps in organizations’ ability to track automated activity and stop agents that behave unexpectedly.
- By Chelsey Arries
- Oct 14, 2026
AI agents may have taken actions that were unintended, unexpected or unauthorized at organizations represented by 49.8% of respondents, who confirmed or suspected such incidents during the previous 12 months, according to Salt Security research.
Among those respondents, more than half reported an unauthorized external application programming interface call or transaction. More than 40% reported changes to or deletion of data.
The company’s State of Agentic AI and API Security 2H 2026 report analyzed more than 380 responses from security and technology professionals alongside operational data from hundreds of customer environments.
While 86.4% reported AI agents operating within their organizations, just 12.5% could reliably follow their activity from the original instruction through connected tools to the result. More than 60% said reconstructing an agent’s previous-day activity would require hours or days of manual investigation.
Only 17.5% reported being able to automatically halt an agent behaving unexpectedly in a live environment. Most could rely only on manual intervention or shutting down the application.
About the Author
Chelsey Arries is assistant editor of Security Today, Campus Security Today, Occupational Health & Safety and Environmental Protection Online.