Cequence and KPMG report on securing autonomous AI agents
As AI agents move from chatting to taking actions like moving money, new research highlights the urgent need for real-time runtime governance.
The short version
- Cequence tracked more than a quarter million AI agent visits across five industries in a single week in October 2026.
- About one in seven agents were found to be booking, buying, signing in, or managing accounts for customers.
- At one telecommunications carrier, agent traffic that failed to identify itself grew from about 100 requests a day to as many as 32,000 within three weeks of a major consumer AI agent’s launch.
- A study showed that agents with defined jobs and real-time checks completed 95% of tasks while using 35% fewer tokens.
- The report recommends 'runtime governance' to check what an agent does at every step rather than just verifying its identity once.
Why it matters
As AI moves from a conversational tool to an autonomous worker, businesses need ways to ensure these agents don't exceed their authority. This research provides a blueprint for maintaining control without blocking the efficiency that AI agents provide.
The shift to autonomous action
Artificial intelligence is transitioning from a tool that simply answers questions to a system that performs consequential actions. According to a report by Cequence Security and KPMG, these AI agents are now capable of issuing refunds, updating records, and moving money on behalf of users. This shift creates a new landscape for how organizations must manage digital interactions.
The report highlights that agents are already active in the real world. During a one-week period in October 2026, Cequence's research team tracked more than a quarter million AI agent visits across five different industries. The data shows that approximately one in seven of these agents were actively booking, buying, or managing accounts for their human customers.
This evolution means that traditional security measures may no longer be enough. Because agents can carry out a sequence of actions, a single permission granted at the start of a task might not be sufficient to prevent a problem at the end. Organizations now need to consider the cumulative risk created across an entire automated workflow rather than just checking a single login.
Hidden agents and security risks
One of the primary challenges identified in the report is that many AI agents do not identify themselves as such. At a Tier 1 telecommunications carrier, traffic from agents that failed to identify themselves grew from about 100 requests a day to as many as 32,000 within just three weeks of a major consumer AI agent launch. These agents were able to sign in as customers and successfully pass multi-factor authentication.
The lack of transparency also opens the door for malicious actors to exploit trusted names. The research found that at one online marketplace, visits claiming to be a 'Google-Agent' outnumbered all genuine agent visits tracked that week. These visits went straight to sign-in pages and failed, suggesting that attackers were testing stolen passwords in bulk while hiding behind a legitimate agent's identity.
These findings underscore the need for better visibility into who—or what—is interacting with a company's infrastructure. When agents can mimic human behavior and bypass standard identity checks, companies must find new ways to distinguish between a helpful tool and a potential threat. The report emphasizes that a trusted name does not guarantee a safe interaction.
Implementing runtime governance
To manage these risks, Cequence and KPMG recommend a strategy of runtime governance. This means checking what an agent does at every individual step of a process rather than just verifying its identity at the beginning. Shreyans Mehta, CTO of Cequence Security, noted that 'an approval at step one shouldn’t carry into step five.'
The report suggests that every agent should be given a specific job description translated into hard technical limits. Instead of giving an agent open-ended authority, organizations should define exactly what it can and cannot do. If an agent attempts an action outside of those limits, a separate control system should block it and point the agent toward the allowed path to complete the work.
Testing this approach showed significant improvements in both security and efficiency. In a study using a Cequence research prototype, agents with only basic logins tried to act outside their jobs in 44% of runs. However, when agents were given defined jobs and every action was checked against those limits, those unauthorized attempts fell to just 1%.
Efficiency and business outcomes
Beyond security, proper governance actually helps AI agents perform their jobs more effectively. In the study mentioned above, task completion rates rose from 76% to 95% when agents were given clear boundaries and real-time checks. Furthermore, these agents used about 35% fewer tokens per run, which can help reduce the costs associated with running large language models.
The goal for organizations is to create a system where good governance allows agents to do more legitimate work without losing control. By providing clear instructions and constant monitoring, companies can ensure that agents stay on track. This approach prevents the agent from wandering into unauthorized territory while still allowing it to complete the tasks the customer intended for it to do.
This framework is particularly relevant for the Asia-Pacific region, where companies are moving quickly from AI experimentation to real-world deployment. As boards and regulators look for ways to manage risk, these findings provide a way to enable legitimate work to continue. The report aims to help leaders put appropriate oversight around AI agents while still capturing the business value they offer.
What to watch
- How telecommunications and retail industries will update their multi-factor authentication to distinguish between humans and agents.
- The development of standardized 'job descriptions' for AI agents to ensure consistent governance across different platforms.
Try it
Review your company's current AI permissions to see if they are 'open-ended' or if they have specific, hard limits on what actions an agent can take.
Sources and links
First reported by Thailand Business News. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.