Summary
- Immersive has introduced AI Govern & Control and AI Observe & Assess, with early access open and general availability planned for early 2027.
- The first product identifies workforce AI usage and sensitive information risks, while the second maps and tests internally developed AI agents.
- Findings can inform targeted training through Immersive One and remediation workflows for developers and security teams.
UK cybersecurity company Immersive has opened early access to two products covering employee use of artificial intelligence and the security testing of autonomous agents developed within enterprises.
Immersive introduced AI Govern & Control and AI Observe & Assess on 18 November 2026, extending its existing cyber readiness platform into AI usage monitoring, risk assessment and intervention.
The products are being introduced through an early access programme, with general availability planned for early 2027. Organisations cannot therefore assume that both capabilities are already available through ordinary commercial deployment.
The two offerings address different sources of exposure. AI Govern & Control focuses on how employees interact with external AI applications, while AI Observe & Assess examines systems and agents developed or deployed by an organisation itself.
Workforce controls examine information before submission
Employees increasingly use generative AI services to analyse documents, prepare material and complete routine business tasks, sometimes through applications that have not been approved or recorded by their employer.
Those interactions can involve sensitive information, particularly when staff copy internal documents, customer records or commercially confidential material into an external service.
AI Govern & Control is intended to identify the AI applications used by employees and assess information submitted through supported browser interactions.
Immersive says the product performs sensitive information classification locally on the employee's device. Raw prompt content does not need to leave the browser for that classification process, reducing the additional data exposure that could arise from sending it to another external analysis service.
Where an interaction presents a risk, the product can provide a contextual warning or intervention before the information is submitted.
Such warnings are intended to allow employees to correct an error without necessarily preventing all use of the application.
The distinction between warning and enforcement remains important. An organisation may permit some information to be shared after an acknowledgement while requiring stricter controls for particular data categories or applications.
Actual protection depends on browser coverage, the classification rules in use and whether employees conduct the relevant activity through managed environments.
Agent assessments extend into connected tools
AI Observe & Assess addresses internally developed agents that use models and software tools to carry out tasks.
Unlike an AI application responding only with text, an agent may be able to retrieve information, interact with application programming interfaces or initiate operations through connected systems.
The consequences of an unsafe instruction depend partly on the tools and permissions available to that agent.
Immersive's product is designed to map the relationships between agents, models, connected tools and supporting services. That information can help identify which systems an agent could reach during normal operation or following manipulation.
The product also supports adversarial testing intended to identify weaknesses before agents are deployed into sensitive environments.
Assessments can examine whether an agent responds appropriately to hostile instructions, respects access restrictions and avoids using connected tools in ways that conflict with its intended purpose.
Immersive says this testing can operate within the customer's infrastructure, allowing sensitive agent information to remain under the customer's control during supported assessment workflows.
Findings can be directed into engineering tools such as Jira and Slack, connecting identified weaknesses with the teams responsible for remediation.
The company has not provided independent evidence establishing that its assessments identify every possible failure mode or that an agent passing a particular test will remain secure after changes to its model, instructions or connected services.
Training forms part of the security process
Both products connect with Immersive One, the company's existing training and cyber readiness platform.
When workforce AI monitoring identifies an unsafe interaction, the finding can inform an exercise addressing the relevant behaviour.
Similarly, agent assessment results may reveal weaknesses arising from development decisions, including excessive permissions or inadequate handling of untrusted instructions.
These findings can inform subsequent development and testing work rather than remaining isolated security reports.
The relationship between detection and training has limits. Employees may understand the risks of sharing confidential information but still make mistakes, while a developer may correctly configure an agent that later acquires access to an additional service.
Technical controls and repeated assessment remain necessary as applications and operating conditions change.
Early access leaves deployment evidence limited
AI governance is becoming a practical challenge for organisations using both commercial AI applications and internally developed agents.
Data protection, access management and software security responsibilities overlap when an employee or agent can move information between systems operated by different providers.
In Europe, the applicable obligations depend on the service, information and processing involved. A monitoring product cannot by itself establish compliance with the EU AI Act or data protection law.
Immersive's offering combines two forms of assessment that are often handled separately: monitoring information sent to AI applications and testing agents capable of taking action through connected tools.
The early access programme will provide a period for testing those capabilities before the planned wider release in early 2027. Independent measures of effectiveness, customer adoption and production performance have not yet been established in the launch material.





