Within a few days, three reports have appeared that converge on the same question: where do AI assistants get their context from, and who controls what flows into it? For companies integrating AI into existing workflows, this is not an academic debate but a question of permissions, data storage and review processes.

Computer History: everyday work as a searchable timeline

OpenAI has introduced a feature called Computer History. On the Mac, it records clicks, keystrokes and app switches and turns them into a searchable timeline that ChatGPT and Codex – OpenAI's tool for programming tasks – can access. The benefit is obvious: an assistant that knows which ticket, file and tool were last worked on has to ask fewer questions and delivers more fitting answers.

The technical implementation is noteworthy. According to the report, the recordings are stored locally as unencrypted Markdown files. According to OpenAI, they are not used for training models. Together, both give a mixed picture: the data does not leave the device automatically, but it is stored there in a format that any person and any process with file access can read in plain text.

For companies, this means concretely: such a timeline potentially contains password entries, customer names, contract contents, internal document titles and fragments from systems for which clear access rules exist. Anyone wishing to use the feature should clarify in advance which devices it may be active on at all, how the files are handled in backups and endpoint management, and which roles within the company can access them.

Prompt injections in the courtroom

How easily AI-supported review processes can be undermined is shown by a case from Connecticut. A plaintiff had embedded invisible instructions in court filings: in 3-point font and white text on a white background. For human readers, the text was practically imperceptible, whereas an automated system could easily have interpreted it as an instruction. The aim was apparently to influence a possible AI review system of the court.

Such manipulations are called prompt injection: attackers smuggle instructions into content that a language model processes, and the model treats them as legitimate user specifications. The responsible judge Spader compared the attempt to secretly influencing jurors and revoked the plaintiff's right to electronic filing.

The incident is also instructive because it did not require sophisticated technology. Font size and font colour in an ordinary document were enough. Precisely this pattern can be transferred to many business processes in which documents arrive from outside and are machine pre-sorted or assessed – applications, offers, complaints, invoices, funding applications.

Consequences for automated review processes

Some practical rules can be derived from the case that apply regardless of the model used:

  • Incoming content is data, not instructions. System specifications and external documents must remain clearly separated in the prompt so that text from a PDF file cannot take on the role of an instruction.
  • Normalise text before processing. Formatting such as tiny font sizes, white text or invisible characters should be noticed and logged during extraction instead of silently ending up in the context window.
  • No consequential decisions without human approval. Where an AI assessment co-decides on rejection, prioritisation or payment, a traceable control instance is needed.
  • Make anomalies visible. A review step that reports manipulation attempts as a finding is more valuable than one that merely ignores them.

Watermarking: detection with clear limits

The third report concerns the opposite direction – the question of whether a text originates from a model. Anthropic intends to provide a verification API for third parties with which texts generated with Claude can be detected more reliably. The method is based on Google's SynthID method and alters the random source in word selection. According to the company, text quality is not supposed to suffer as a result.

Equally important are the stated limits: with strongly fact-based texts, with code and with heavy rewriting, the method can reach its limits. That is plausible, because the watermark lives on leeway in word selection. Where there are hardly any alternatives or where a text is subsequently reworded, the signal disappears.

For practice, this leads to a sober expectation. A watermark can be one indication among several, for example in editorial processes or when reviewing submitted content. It is not evidence and, moreover, only covers texts from the respective provider. Processes that rely on a clear yes-no answer will not work reliably with this technology. The timing of availability is also still open; what has been announced is the provision, not the launch.

The common denominator: approvals instead of trust

All three developments show the same shift. AI assistants become more useful the more context they receive – and more vulnerable the less it is checked where this context comes from. A timeline of everyday work is a productive knowledge store and at the same time a sensitive data set. An automated document check saves time and at the same time opens up a new attack surface.

A watermark creates traceability, but only within a limited scope.

Anyone integrating AI into workflows should therefore answer two questions early on: which data sources may an assistant see, and at which points does a human decide?Both belong in the architecture, not in a retrospective policy.

For web and software projects, this means: context accesses and approval levels belong in the permission concept from the outset, with documented data sources and a clearly defined point at which a person decides.Every pipeline that processes third-party documents needs a pre-check for hidden content and a log of what went into the model.And detection methods such as watermarks are useful as an additional indication, not as the basis for automatic decisions.

Sources