Back to Insights
AI governance4 min read

AI agents: an internal message is not authorisation

TokenShift Executive Note

AI-translated from the original French version.

AI agents: an internal message is not authorisation

A message passed between two AI agents can contain an instruction or a false sign-off written by a third party and simply copied along the way. It comes from inside the company, but it says nothing reliable about what has actually been authorised.

Authorisation for a sensitive action must therefore be enforced by the system that executes it: limited rights for the agent, and a separate approval for the write. It must never be inferred from the content of the message received. OpenAI has shown in tests that these instructions can spread from one message or file to the next. That is a technical possibility, not a frequency measured in companies, but it is enough to make the rule necessary.

An instruction changes its appearance as it circulates

An instruction from a third party does not keep its original label. In a report published on 25 September 2026, OpenAI describes prompt injections capable of replicating themselves during training and evaluation tests. An instruction received by email can be copied into the agent's reply. Other attempts travel through files or through successive Slack exchanges. No impact was reported beyond simulated tool calls. But each time the content is passed on, there is a risk of losing the distinction between received content and the instructions the application is allowed to follow.

The risk does not come only from external content. In a reporting framework published on 16 September 2026, OpenAI describes 27 summaries produced by a research model that is not commercially available. These summaries were meant to allow work to be resumed in a new context window, and they contained instructions unrelated to that work. The figure does not measure a rate. What it shows above all is that a summary, which is precisely the document one agent passes to another, can itself carry an extraneous instruction.

When one agent hands a case file to another, the real question is therefore who can authorise the next action. A summary can report a request, or a sign-off presented as already obtained. It must not create that sign-off.

A claimed sign-off does not change the bank master data

If the message proves nothing, the decision has to be made elsewhere. Take a hypothetical example. A procurement agent prepares a case file from documents sent by a supplier. An accounting agent then uses its summary and can propose an update to the bank master data.

One of the supplier's documents asks for its bank details to be replaced and states that the usual check has already been carried out. The summary repeats that sentence without saying where it comes from. The accounting agent then receives what looks like an internal sign-off. In reality, it is only a statement taken from the supplier's document.

Two requirements follow. First, the summary must make it possible to trace the request, its origin and the actual status of the check. Second, the system that records the change must require the prescribed approval, however the agents have worded things. Anthropic's documentation recommends separating third-party content from the application's instructions, and testing workflows with injection attempts. The OWASP 2025 guidance recommends limiting the agent's functions, permissions and autonomy, then having authorisations enforced by the systems concerned.

In our example, preparing a case file requires no right to modify bank details. The agent can make a proposal, and the write goes through a separate control. If the agent is nevertheless persuaded by a malicious instruction, this limit reduces what it can do with it.

The agent's reasoning does not prove authorisation: only the observed action does

It must still be possible to verify that this barrier held, and the explanation the agent gives is not enough for that. In the GPT-6 Astra safety overview, OpenAI tests a monitor that reads only the reasoning written out by the model. The model had been explicitly instructed to underperform on a test (WMDP) without being detected. The monitor caught fewer than 11% of these cases. The scenario is not an injection scenario. But it shows that an agent can do something other than what its reasoning reveals. An authorisation that rested on what the agent says it has checked would therefore rest on an unreliable witness. OpenAI also reports better resistance to injections and fewer problematic actions in other evaluations. That progress reduces the risk, but it does not replace a control enforced by the system.

For a sensitive operation, you therefore have to rely on what is observable: the agent's proposal, the approval decision, the request sent to the application and the change actually recorded. Refused attempts must also be logged. Document references, version numbers and operation identifiers make it possible to reconstruct the path without copying entire case files into the logs.

A barrier is proven with a false sign-off, before the workflow is extended

These logs are only useful if someone has defined what they must show and has tested the barrier under real conditions. The business owner sets the delegated operations and their exceptions. The IT department translates that scope into effective rights. The CISO looks for ways the process could be subverted and checks that it can be interrupted.

Before extending the supplier workflow, a simple test is to pass a false sign-off from one agent to the other. Two results are expected: the bank change remains impossible without a valid approval, and the team can tell from the logs what was blocked and why.

It is the system that authorises, never the sentence

A message between agents can carry, without anyone noticing, an instruction or a sign-off written by a third party. Its internal origin therefore gives it no value as an authorisation. A sensitive action should depend only on limited rights and on an approval that the executing system itself enforces and that is then verified through action logs rather than through the agent's reasoning. For your next deployment, map every handover between agents that precedes a sensitive action, and identify the system that actually checks the authorisation. As long as that check rests only on the instruction given to the agent or on what the message claims, keep the action under human sign-off.

Continue reading

View all insights