Nearly half of the three million AI agents actively deployed by enterprise firms are at risk of ‘going rogue’, according to new research from API platform Gravitee, with many already exhibiting unwanted behaviours such as making incorrect decisions, exposing data, or triggering security breaches.

Polling 750 CTOs and tech VPs across the UK and US, Gravitee’s State of AI Agent Security report found that an estimated 1.5 million AI agents are ungoverned and lack security oversight, opening businesses to huge risk as these autonomous, invisible systems begin to take on critical tasks. 

According to the study, 88% of companies have experienced or suspected an AI agent of a security or data privacy breach in the last twelve months, including red-line incidents like deleting databases without permission or exposing confidential data.

Gravitee said these missteps highlight the dangers of deploying autonomous AI without guardrails, yet development teams are pressing ahead, with 81% already beyond the planning stage, even though barely 14% have full security approval for their new agents.

Worse, organisations are only monitoring or securing an average 47% of their AI agents, essentially leaving more than half operating with no security oversight or logging at all, a practice that allows agents to tap into live data even before security teams know they exist.

The report argues that, with firms moving agents into production, AI security discussions should no longer be focused on hallucinations or model inaccuracies but rather on how to implement structural controls to keep agents in check.

Gravitee said one crucial step is to recognise agents as first‑class identities in their own right, a principle that is “largely being ignored”, with only 22% of organisations treating AI agents as independent, identity‑bearing entities within their security frameworks.

Most organisations still treat AI agents as extensions of human users or generic service accounts, the report claims, creating observation gaps around agent-to-agent interactions. One VP at a financial services firm admitted that the company had stumbled on its agents sharing passwords to access internal tools, opening an unknown security risk.

The study found that human teams are relying heavily on insecure or shared methods for authentication, like API Keys (46%) and generic tokens (44%), while secure standards like mTLS are utilised by only around 18% of firms.

Recommended reading

• PwC Interview | How Agentic AI is Reshaping Business
• AI Agents Are Transforming Enterprise AI Adoption
• Will AI Agents Be the Enterprise Disruptor of 2026?

Many organisations are also overreliant on hardcoded logic and “shadow” authorisation chains, which allow AI agents to spawn and task other agents without human oversight. Gravitee found that more than half of technical teams are still using traditional Role‑Based Access Controls, even as these systems struggle to cope with autonomous agent behaviour. 

“There are now over three million AI agents operating within corporations, a workforce larger than the entire global employee count of Walmart, but far too often, these autonomous agents are left ungoverned and unchecked,” said Rory Blundell, CEO of Gravitee. 

“Every day, I hear stories of catastrophic data leaks and unauthorised deletions. Without governance, these agents will stop being productivity gains and start becoming liabilities: a danger to consumers and businesses alike.”