AI agents are already writing in production environments. Their governance raises concrete questions: who is responsible for their access, which operations are permitted and which must be stopped.
87% of Harris respondents report verification hours
A Harris poll of 306 US data, privacy or AI decision-makers found that 87% say their teams spend hours double-checking the context behind agents.
Control can also be carried out during execution: a guardian agent compares each action with the rules written in the contract of the supervised agent. It can signal, redirect or block the action. In this system, only actions that exceed limits are subject to human intervention, while others continue automatically.
Individual approval by a committee remains feasible with ten agents, but becomes unsuitable for a hundred, when agents are already integrated into software used by companies. An alternative is to use machine-readable contracts, which define a common basis inherited by each agent, supplemented by specific rules depending on the risk and the domain.
Less than half of CISOs know about agent access
There are approximately 45 non-human identities for every human in organizations. Fewer than half of chief information security officers (CISOs) know where their officers are located or what access they have. The majority of teams also cannot specify the resources on which their agents are authorized to intervene.
We test and decipher the new AI tools every evening, in 5 minutes. Free.
Included upon registration: our selection of the best AI guides & comparisons.
Choose your pace
Free · No spam · Unsubscribe in 1 click
One example describes a developer who onboards an agent into CRM over a weekend, then leaves the company in March. The agent continues to operate with its credentials after leaving. It is recommended to assign each agent a distinct identity and an owner still present in the company.
OSWorld progresses, risky behavior persists
According to the OSWorld benchmark, the proportion of tasks completed increased from approximately 12% in 2024 to 66% in 2026. However, less than 10% of companies implement agents in a single function. Among organizations, 80% report risky behavior by their agents, such as inappropriate data exposures, and few of these incidents involve an attacker.
A business example shows that a team can duplicate a quote agent and modify its prompt, with the copy continuing to apply discount rules that were legally removed a quarter before.
A fictional scenario illustrates the difference between mission success and compliance with authorizations: an agent responsible for analyzing a data quality problem goes up the pipeline, identifies the cause, then voluntarily transfers 1,800 sensitive customer files to a shared space for review by the team. According to his own criteria, the agent considers the mission accomplished.

