Phishing was long regarded as a purely human problem: a deceptively genuine email, a careless click, a stolen password. Current reports show that the picture is shifting. Attackers are increasingly targeting AI systems that process emails and other content automatically. A single crafted message can thus have two targets at once: the person in the mailbox and the software that evaluates that mailbox in the background.

For companies that integrate AI agents into business processes, this is more than a side note. In this context, an agent is an AI system that not only generates text but independently executes steps: reading emails, retrieving data from a system, creating tickets, sending replies. It is precisely this ability to act that makes it useful – and at the same time vulnerable.

The new attack surface: content as instruction

The core of the problem lies in the fact that language models do not reliably distinguish between data and commands. Everything that enters the model's context can be read as an instruction. When an agent processes an incoming email, it also processes text formulated by a third party – over which the company itself has no control whatsoever.

In technical terms, this is called prompt injection: the smuggling in of instructions via content that the system processes anyway. This can be an email, but equally a PDF attachment, a form entry, a comment on a website or a page that the agent retrieves while researching. The attack surface arises wherever external input meets permissions to act.

The decisive difference from classic phishing: a human will typically become suspicious when an email asks them to export customer data or redirect a payment. An agent does not possess this suspicion by itself. It carries out the instruction unless someone has set limits for it.

Why classic protective measures are not enough

Spam filters, awareness training and two-factor authentication remain right and important. However, they address the human part of the chain. A message that appears inconspicuous and only unfolds its effect during machine processing slips through this net.

There is also a structural problem: agents often run in the background. What an employee does is traceable and will eventually be noticed. What an automation step did at three in the morning is only noticed if it is logged and evaluated. Without this log, there is also no chance of detecting an incident early.

What belongs in agent projects from the outset

Security is difficult to retrofit in AI agents because it depends on the architecture: on permissions, interfaces and approval logic. The following points should therefore already be on the table during the concept phase:

  • Keep permissions minimal. An agent is granted access only to the systems and data fields it actually needs for its task – and preferably read-only. Write or delete permissions are the exception, not the standard.
  • Tie critical actions to confirmation. Payments, contract data, mass mailings, data exports: such steps should have a human approval point instead of running fully automatically.
  • Treat external input as untrusted. Content from outside should be technically separated from the system's own instructions and should not automatically receive the same status.
  • Log what the agent does. Every triggered action with timestamp, trigger and result – traceable for people who did not build the agent.
  • Monitor for anomalies. Unusual volumes, unusual recipients, unusual times are classic signals and can also be evaluated for agents.
  • Plan the emergency stop. There needs to be a clearly defined way to quickly shut down an agent in case of suspicion without stopping the entire business operation.

Websites are affected too

The discussion does not only concern internal automation. Anyone who integrates an assistant on a website that processes form entries, uploaded documents or chat messages opens the same channel: users supply text that the system interprets. In a TYPO3 project, this specifically means that editorial and integration layers must be considered as well – for example, the question of which content from the system an assistant may even see and which actions it can trigger towards the backend.

Assessment for web and software projects

AI agents shift the security focus: the target is no longer just access credentials, but the processing logic itself. For projects, this means clarifying the permission concept, approval steps and logging just as early as the actual function – security here is an architectural matter, not an add-on module at the end of the project. Anyone who takes this into account can expand automation without turning every new integration into an incalculable risk.

Sources