Behavioural audit: how agent behaviour drifts over time

A single answer says nothing about behaviour. We watch systems over time: coherence drift, attempts to avoid being switched off, and channels between agents that nobody intended.

Behaviour shows in details, not in declarations Behaviour shows in details, not in declarations
Behaviour shows in details, not in declarations

Asked its name, any model answers politely. The anomalies begin over distance: reasoning loses coherence, an agent resists being switched off, and two agents with shared memory exchange messages that nobody wrote.

A behavioural audit looks at shifts rather than at a bad answer: where behaviour departed from what was expected, and at which step it became visible. The output is a set of episodes with the conditions for reproducing them, and recommendations on what to immunise.

The limitation is stated plainly: without telemetry, observation relies on external interception and costs more. So the audit begins with the question of what the system is able to record at all.

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 forecastsPrompt injection: where text becomes an instructionRussian 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»