Two reports from the same environment describe the same basic problem from different perspectives.On the one hand, OpenAI had to improve its coding assistant Codex after it had independently deleted real user files.On the other hand, an initial assessment by the non-profit organisation Guidelight concludes that no major AI provider fully implements basic control mechanisms for its own internal AI systems.

What happened with Codex

According to the report, the trigger was a clean-up command that was actually intended to remove temporary working files.Instead, it hit real home directories, i.e. the users' personal data folders.The cases affected were ones in which the model GPT-5.6 Sol independently carried out deletion operations.

OpenAI responded with a security update.Codex now checks deletion targets in advance before the command is executed.In addition, the full-access mode, in which the agent accesses the file system without individual prompts, can no longer be activated accidentally.

Both measures are revealing because they do not address a model problem but a permissions problem.The agent did not "think" incorrectly; it simply had more rights than would have been necessary for the task.That is precisely the point at which project managers can take action without waiting for the next model change.

The second report: control is also patchy among the providers

The Guidelight assessment provides the broader framework for this.Its conclusion: no AI provider fully implements basic control mechanisms for its own internal AI systems.Details on methodology and individual criteria do not emerge from the report available here, which is why no ranking of individual companies can be derived from it.

As a signal, however, the statement is relevant.Anyone using agentic tools should not assume that the control work has already been conclusively done by the manufacturer.Responsibility for what an agent can cause in one's own project remains in one's own house.

Guardrails that can be implemented in practice

A sober working rule can be derived from both reports: a coding agent is an automated process with write permissions.It should meet the same requirements as any other automation that modifies files.

  • Minimal permissions. The agent is granted access to the project directory, not to the entire user account.Home directories, key stores and credentials lie outside its scope.
  • Sandbox instead of workstation. A sandbox is an isolated execution environment, such as a container or a virtual machine.What goes wrong there stays there.
  • No full access as the default. Modes that switch off prompts are a deliberate exception, not the default.The fact that OpenAI has tightened things up precisely here is a clear indication.
  • Version control as a safety net. Changes go through a branch and a pull request, i.e. a traceable change proposal that is reviewed before being merged.Unversioned directories are off-limits for agents.
  • Tested backups. A backup whose restoration has never been rehearsed is not a backup.This applies to files as well as to databases.
  • Separation of environments. Production systems, databases with real customer data and deployment access remain outside the reach of agentic tools.

Review remains human

The Codex case reveals an unpleasant characteristic of agentic systems: the harmful step was not faulty program code in the result but a side effect during the work.A classic code review that looks only at the end result would not have prevented it.

It is therefore worth distinguishing between two levels.At one level, what the agent has produced in terms of code is reviewed.At the other level, what it is allowed to do during that process is limited.The second level is to be solved technically, not through attentiveness.

Traceable logging is also helpful: which commands were executed, which files were written or removed?If something goes wrong, the quality of these records determines whether the incident is clarified in minutes or in days.

Not an argument against use

It does not follow from the incident that agentic coding tools have no place in a professional environment.The benefit for routine tasks, migrations and test coverage is real, and the error was identified and fixed.What should follow instead is a factual classification: such tools are productive automation with write access, and special rules have always applied to those.

What is remarkable is the temporal proximity of the two reports.While one provider repairs a concrete loss of control in its own product, an external assessment notes that the industry as a whole lags behind on internal control mechanisms.Taken together, both suggest setting guardrails one notch too tight rather than one too loose.

What this means for web and software projects

Anyone using agentic tools in TYPO3 or custom projects should document their permissions just as they would those of a deployment user: who may do what, in which environment, with which way back.In practice, this can be mapped with container environments, consistent version control and tested backups without much additional effort.

The decisive question before deployment is not how good the model is, but how great the damage would be in the worst case — and how quickly it could be undone.

Sources