Singapore – Cequence Security has published a report, developed with input from KPMG in Singapore, finding that AI agents are already signing in, booking and buying on behalf of customers, often without identifying themselves.
The report examines how organisations can govern AI agents as they move from generating content to taking consequential actions. Those actions now include issuing refunds, updating records and moving money.
To measure the scale, the research team tracked more than a quarter of a million AI agent visits across five industries during one week in October 2026. About one in seven of those visits booked, bought, signed in or managed an account for a customer.
Many of those agents did not announce themselves. At a Tier 1 telecommunications carrier, traffic from unidentified agents grew from about 100 requests a day to as many as 32,000 within three weeks of a major consumer AI agent’s launch.
Those agents signed in as the customer and passed multi-factor authentication.
The data also showed that a trusted name offers little assurance. At one online marketplace, more visits claimed to be Google-Agent than all the genuine agent visits tracked that week.
Every one of those visits went straight to the sign-in page and failed, which suggests attackers were testing stolen passwords in bulk under a legitimate agent’s name. Google was not involved.
Against this backdrop, the report recommends giving each agent a specific job rather than open-ended authority. Each agent should work from a plain-language job description converted into hard limits, so it cannot decide on its own to do more.
It also calls for a separate control outside the agent to approve, escalate or block every action in real time, because harmless-looking steps can add up to something the agent is not allowed to do. When that control blocks an action, it should point the agent to the permitted route so the work still gets done.
In addition, the report argues that success should be measured by work completed rather than actions blocked.
A controlled study using a research prototype put numbers on the approach. Agents with only logins and fixed permissions tried to act outside their job in 44% of runs.
Once each agent had a defined job and every action was checked against it, those attempts fell to 1%. Task completion rose from 76% to 95%, and agents used about 35% fewer tokens per run.
“Organisations across Asia-Pacific are moving from experimenting with AI to deploying them to achieve real business outcomes. Traditional controls such as identity verification and fixed access permissions are important, but they may not be sufficient when an agent can carry out a sequence of actions on behalf of a business or customer,” stated Gerry Chng, Partner, Head of Cyber, KPMG in Singapore.
Organisations need governance that considers the agent’s role, the context of each action, and the cumulative risk created across an entire task. KPMG’s contribution to this report brings together our experience in governance, risk, regulation and cyber security to help organisations put appropriate oversight around AI agents while enabling legitimate work to continue,” continued Chng.
Shreyans Mehta, co-founder and CTO of Cequence Security, who also co-authored the report, framed the issue in terms of continuous checking.
“An approval at step one shouldn’t carry into step five. Runtime governance means checking what an agent does at every step. When the answer is no, it means telling the agent why, so it gets back on track instead of looking for a way around. It has to work the same way for an agent your customer sends as for one you built yourself,” said Mehta.

