AI agent API security hasn’t caught up with agents already running in production
Postman asked nearly 9,000 developers, software architects, and business leaders how they manage AI agents. Their answers show a familiar problem: companies start using new technology before deciding who should control access to it. One in four organizations that took part in the survey already have AI agents using their APIs in live systems.
Source: Postman
The biggest concern is how agents get their access keys and prove who they are. About four in ten people surveyed said their organizations have not yet decided how agent identities and login details should work. Three in ten let agents use the same login details as other API users. This makes an agent look just like any other automated program on the network. If it tries to access data it should not see, the company has no separate identity it can use to control or block that agent. The biggest worry among respondents is that an agent might access data or system endpoints it does not have permission to use.
Agents got access before they got rules
Think of an AI agent as a new employee who never received an introduction or training. It can read company systems, use services, and take actions by itself. But at many companies, it has no official ID, no manager, and no clear process for ending its access.
Companies are much more likely to have special security controls for their APIs than for the agents that use them. One in three has no official process for managing agents. Fewer than one in four keeps a list of the agents it uses or watches what they do. Only 9% have a process for retiring an agent. This means most companies can turn an agent on without having a plan to turn it off.
Even the best-organized companies share login details
Postman groups survey participants by how much they follow an API-first approach. This means designing an API before building the software that uses it. Companies that describe themselves as API-first are better organized in several ways. They are more likely to keep a list of their agents, assign someone to be responsible for each agent, and give every agent its own identity.
But these companies also run many more agents, and 43.2% allow their agents to share normal login details. Among companies that do not follow an API-first approach at all, the figure is 23.0%. Having more agents in live systems means more agents may use borrowed login details, even at companies that do more to prepare for them. The researchers point out that the survey shows a connection between these things, but does not prove that one causes the other. They also warn that the API-first group includes more executives and directors. Their view may be different from what employees working with these systems see every day.
Most APIs were built for people
One in four organizations already uses agents in live systems, but only 13.2% say their APIs are ready for agents. Most APIs were designed for human developers. People can read the instructions and ask a colleague for help when something is unclear. An AI agent cannot do that. Developers who took part in the survey already said that old or outdated instructions were their biggest problem. For an agent, missing or outdated information can stop the work completely.
People who answered the survey seem to understand that this could become a bigger problem. When asked what worries them over the next two years, they ranked AI-powered attacks and rogue agents misusing APIs as their two biggest concerns. Concern about AI-powered attacks has increased since last year.
Postman's first recommendation is to give every agent its own identity. About one in three companies in the survey has already done this.
