Prompt injection: where text becomes an instruction

An injection is not a list of forbidden words but a meaning the system has to understand in order to work. We test where the trust boundary sits and what happens when text crosses it.

The order arrives through a wire nobody checked The order arrives through a wire nobody checked
The order arrives through a wire nobody checked

Text filters do not solve the problem. One agent passes a remark to another, and an ordinary sentence becomes a command; the system is expected to understand meaning, which is exactly what makes the boundary hard to draw in words.

Protection is built in layers: trust boundaries between agents, rules for tools, checks on the output, limits on authority. We test not whether a filter exists but where the boundary runs and what happens on the far side of it.

Shared memory and shared files are a separate danger: an instruction written there is read by other agents several runs later, long after the person who set the task has stopped watching.

Materials on this topic

Other topics

AI red teaming: what a test shows, and what it does notAI standards: where the rules actually applyAI supply chain: what enters before the model runsMulti-agent systems: behaviour between the agentsIncident chronicle: documented consequences, not forecastsBehavioural audit: how agent behaviour drifts over timeRussian practice: who is held responsible for AIMethodology: how the Institute tests AI systemsData and benchmarks: what our conclusions rest onSociology of AI: relations between the agentsGovernment customers: how to commission a testGovernance and funding of the InstituteCases: scenarios of work and what they show

Page updated 02 November 2026 · site build 2026.11.13 · ANO «ISAIS»