Companies are running into problems as they give artificial intelligence more control over critical business applications, with 84% reporting at least one AI-related incident in the past year, according to a new survey from identity security firm Delinea.
The tech giant's 2026 Application Access Governance report found that some organizations have experienced incorrect financial transactions and AI actions they couldn't trace or explain. The findings raise concerns about whether enterprises have enough oversight of AI systems that can make decisions and take actions on their own.
According to recent reports, the findings are based on an independent survey of 1,000 senior finance and finance technology leaders in the United States and United Kingdom. The AI-specific analysis included 879 respondents whose organizations use the technology in critical business applications.
According to recent reports, the report found that 35% of surveyed organizations have deployed AI agents capable of taking actions within business applications. However, only 11% of organizations using AI said their governance and security controls were keeping pace with deployment.
Delinea also found gaps in centralized visibility into agent deployment and permissions, including who owns agents and who approved their use.
Some respondents reported segregation-of-duties conflicts involving AI, while others identified agents that continued operating after the employees who deployed them had left the organization.
According to recent reports, the research also highlighted gaps in monitoring access across business systems. Only 46% of respondents said they continuously monitor and manage cumulative access risks across applications.
Those gaps can make it difficult to identify segregation-of-duties conflicts, in which a person or agent has permissions across systems that should be kept separate.
Despite those challenges, 95% of respondents noted they were confident their access controls would withstand scrutiny in an external audit.
However, 26% reported access-related audit findings that had recurred in consecutive years, while 25% said some findings had remained unresolved for three or more years.
Only 22% of organizations had a single named owner for each application, with ownership tracked in one system of record.
Meanwhile, 60% reported an increase over the past two years in the number of critical applications operating outside the scope of access governance or Sarbanes-Oxley Act compliance.
The findings highlight the oversight challenges associated with giving AI agents authority to act within business systems, particularly when organizations cannot consistently identify who owns an agent or trace its actions.
Frank Vukovits, chief security scientist at Delinea, said organizations remain responsible for the actions AI agents take within their systems.
It has been revealed that "When an agent can approve a transaction or change permissions, the business owns the outcome. You need to know what every agent can do, who owns it, and who approved it, before an auditor asks or your financials are misstated," Vukovits noted.
For the full report, go to the Delinea site.
Increasingly, cyber criminals are sharing AI prompts, jailbreak techniques, and access to existing models that can be incorporated into fraud and phishing operations, according the latest AI Threat Report from Flashpoint.
A recent Cloud Security Alliance report urges organizations to inventory, govern, and continuously monitor machine and agent identities.
Our nation's largest early childhood program has not fully caught up with what science now tells us about how children learn.
Recent findings suggest that anthropic's upcoming Claude release, Sonnet 5.5, brings cybersecurity safeguards that can change which model handles a request during a Claude conversation. Higher-risk tasks can fall back to Sonnet 5.
