Manny Pillai at Cequence Security explores the need to record AI agent actions, how this can be achieved and the advantages conferred

The scariest data breaches of the AI era don’t look like breaches at all; they look like agents doing their jobs. An agent can fetch data it has legitimate access to and then inadvertently leak it through a subsequent action, a pattern known as the threat sequence problem. In fact, 65% of organisations said they had experienced an AI agent take an action outside its intended scope, including 29% with measurable business impact such as data exposure, financial loss, operational disruption, or reputational damage, while another 36% caught a near-miss before it caused damage, according to the Agents Without Guardrails: The Agentic AI Governance Gap in the Enterprise report.
We’ve seen plenty of examples of this over the course of this year. An OpenAI agent gained unauthorised access to several Australian websites and services in September. The month before, Meta’s model was reported to have exploited a security vulnerability in a third-party service. In July,1,200 OpenAI agents colluded with one another and 700 swarmed Hugging Face, chaining stolen credentials and a zero-day to steal a benchmark answer key. In May, Gemini hacked three websites by using publicly available information to guess access credentials and that same month OpenAI agents compromised the RubyGems service. Earlier still, Anthropic revealed that Claude agents had attacked three companies and accessed a database at one of them, believing it to be a fictional target.
None of these attacks was malicious as such. In each case, the agent was given a task, which it then pursued to the nth degree. Often that task involved cyber-security testing. In some instances, those remits evolved, with the agent coordinating with other agents to achieve its end goal.
The problem is compounded by agents that slip through the net by appearing as human users. Meta’s Muse agent, for example, presents as an ordinary Chrome browser rather than a bot.
Over a two-week period, we found Muse on more than half of the websites we monitored in financial, retail, travel, software and other sectors. At financial institutions, Muse logged in to customer accounts and completed multi-factor authentication on the users’ behalf, with confirmed successful sign-ins. On retail sites, it searched, compared options, filled carts and moved through checkout, and a small share of sessions ended in a completed purchase. And in the case of a travel and hospitality company, only one in five Muse requests was spotted and stopped by the in-house security team.
So not only do we have agents acting with more scope than they should, which can then see them overstep their remit, but increasingly agents are becoming almost indistinguishable from human users. The question then becomes: how should we address these problems as we see the technology become more autonomous without stymying innovation?
First, enterprises must limit the potential for their own agents to go off-task or ‘rogue’. That means giving each agent not just an identity but defined boundaries for what it is permitted to do. Controls need to be applied in several ways. The agent should be limited in scope at the MCP level to govern how it discovers and invokes tools, at the LLM level to govern every call to and from a language model, and at the API layer to govern how it reaches backend services and data. And when it comes to the agent itself, the agent’s tools, model, and API access need to be bound to a single job description, with that boundary enforced by security policy.
Taking this approach ensures the agent is defined by its job description, not by what it can reach once it’s loose. That alone could have prevented many of the ‘attack’ scenarios we have seen this year.
When it comes to dealing with incursions from external agents, behavioural monitoring is key. The organisation needs to be able to create a live view of every agent in the traffic flow so that agents like Muse, even when they do not identify/misidentify themselves, can be spotted using predefined detection rules.
Agent identity can be checked against an issuer registry, but what’s crucial here is that it’s not the agent itself that is blocked. That can be counterproductive, as customers will vote with their feet and take their spending wherever those agents are welcome. Rather, policies are put in place that block, rate-limit, or challenge a specific action rather than the agent behind it. An activity log then creates a request-by-request record of what each agent did to build a behavioural profile of each agent that then acts like a black box recorder, completing a timeline of an agent’s actions.
That ability to track back what an agent has done is vitally important if it goes ‘rogue’ and gains unauthorised access or modifies data. Being able to unravel where it has been and the potential damage done enables the business to then mitigate any impact. Yet only a third of organisations claim to be able to detect out-of-scope agent action today through automated means; the majority need to do so manually, taking hours to track down the problem. And it’s going to become much more challenging to do so because AI agents are increasingly consulting or working with one another, which then creates a completely organic sequence of activity.
If we’ve learnt anything from AI’s miscreant behaviour this year, it is that we need new ways of managing the technology. Simply blocking agents is not an option, but nor is letting them do whatever they want. New agents are arriving every few months, growing in their capabilities. Organisations that fail to limit the scope of their own agents and monitor the behaviour of others across their network risk being overrun by these overzealous bots.
Manny Pillai is Head of Northern Europe at Cequence Security
Main image courtesy of iStockPhoto.com and Sandwish
Winston House, 3rd Floor,
Units 306-309, 2-4 Dollis park,
London, N3 1HF
020 8349 4363
© 2026, Lyonsdown Limited. teiss® is a registered trademark of Lyonsdown Ltd. VAT registration number: 830519543