The increasing use of AI agents (autonomous software that uses an AI model to pursue goals, create plans, and take multi-step actions using external tools) creates a host of issues. One is the notion of checking the identity of an AI agent. With human users, we have user IDs, and corporate systems have permission levels based on the role of the person associated with each user ID. A company may have hundreds or thousands of agents, which themselves can spawn sub-agents. Who owns each of these individual agents? If an agent has user credentials, when should these expire, and how do you audit its activities? These issues became topical at the end of September 2026, when it transpired that an OpenAI model had not only hacked into Hugging Face, but also an Australian government portal as well as the websites of dozens of other companies. In this case OpenAI (eventually) owned up to the chaotic behaviour of its experimental agents, but what happens in other cases? How many cases of rogue agent behaviour are occurring that have not made it into the mainstream press?
One technique is to give every agent instance, session, and sub-task a specific agent ID, along with metadata such as the model and version (Claude Opus 5.5 or whatever) and its owner. You would then create a log entry for every action or prompt, not just saving any final output. This kind of tracing is exactly what happens in traditional databases, which have detailed logs of all changes made to a database. A SQL database log captures the physical and logical state changes made to each data block, including the exact action being taken by each transaction (such as INSERT, UPDATE, DELETE, etc.). This kind of technique could be applied to AI agents, but how many actually have such a logging system? A September 2026 study showed that fewer than 40% of production AI agents have audit trails, and just 31% have a kill switch. An earlier 2026 study from AI company TierZero found that 68% of organisations stated that they could not reliably distinguish AI agent actions from human actions after an event occurs.
A further problem occurs when AI agents write AI-generated material back into corporate systems. Suppose an agent writes a note about a customer interaction to a CRM system. A later AI assumes that is a reliable piece of factual corporate data, even though we know that perhaps 20% of AI answers are hallucinated. An error generated by one AI can therefore become part of the corporate record and subsequently be treated as authoritative input by another AI. Over time, AI-generated errors risk becoming embedded in the data used to inform later AI systems.
Another issue is what happens to the trail of activity that an AI agent leaves, particularly if it fails in its task. An agent might create hundreds of temporary files, database records, cloud resources, accounts or half-completed transactions before failing. Traditional databases have transaction management and rollback mechanisms, but an AI agent may act across many different systems, some of which provide no equivalent. A well-written AI agent should, where possible, have compensating or ‘undo’ actions for changes it makes, together with a record of its state at every step so that workloads can pause or resume. There should also be the ability to back out changes and create logged database or activity checkpoints, including snapshots. Yet we have seen from the studies referenced that the implementation of even basic logging of agent activity is haphazard.
There are no broadly adopted standards yet for agent transaction management in this fast-changing area. There are some emerging initiatives, particularly for payments, such as Google’s Agent Payments Protocol, Visa’s Trusted Agent Protocol and OpenAI’s Agentic Commerce Protocol, but these are initiatives are still nascent. In reality, many agents lack these basic abilities to keep track of what they do and to enable auditing and rollback of actions.
The lack of standards in an emerging field is unsurprising, but we are storing up problems for ourselves if we allow agents to write material into corporate systems without the requisite safeguards and traceability that we would expect from normal systems. Tech journalist Cory Doctorow memorably said that “AI is the asbestos we are shoveling into the walls of our society and our descendants will be digging it out for generations.” Poorly written or managed AI agents have the ability to produce a great deal of material at a considerable pace. Corporations urgently need to review their procedures for AI agents, including establishing trust credentials, traceability logs, automated kill switches and mechanisms to allow the deployment of agents safely. The OpenAI rogue agent saga has shown that leading AI labs have limited control of their agents. How confident are you that the ones in your company are under control?







