[Digital Today reporter Chi-gyu Hwang (황치규)] Darktrace has formally launched a security product called "Secure AI" to inspect the use of AI tools and AI agents inside companies. SiliconANGLE reported on Sept. 22 that the product focuses on tracking and controlling corporate AI use that is spreading faster than security teams can manage.

Darktrace expanded its existing behaviour-based detection approach to AI environments. It learns what normal activity looks like at each customer and then looks for anomalies.

Darktrace said telemetry data collected from about 8,200 deployments showed that, as of August, more than 80 percent of monitored customer accounts used generative AI services. Organisations used an average of 5 AI providers that month, it said.

Early deployments also identified uncontrolled AI use. One organisation approved only 1 AI assistant, but most employees used unauthorised services. Another organisation created about 90 AI agents in a low-code development environment without an approval process.

"Secure AI" monitors ChatGPT Enterprise, Claude, Microsoft Copilot and Amazon Bedrock sessions in real time. It checks for jailbreak attempts, exposure of sensitive information and signs of indirect prompt injection, and assigns a risk score to each session so security staff can prioritise responses.

For shadow AI detection, it also uses integration with secure access service edge offerings such as Microsoft Entra Global Secure Access.

Administrators can block unauthorised services or isolate devices. Companies can also upload their own AI policies as free-form text and check them against regulatory and compliance frameworks.

Darktrace links AI agents to actual users for tracking and records each agent's access scope. It also monitors development environments such as Amazon Bedrock and Microsoft Copilot Studio so it can flag excessive privileges or configuration errors during the agent build stage.