Darktrace
Darktrace Launches Tool to Govern Shadow AI and Agent Risks
Darktrace has launched Darktrace / SECURE AI, a product designed to monitor and govern artificial-intelligence use across organisations. The company says more than 80% of its customers used generative AI services in August 2026, based on telemetry from approximately 8,200 observed deployments, with the average organisation using five AI providers.
Darktrace says conventional, rules-based controls can struggle to judge whether prompts, agent permissions and combined actions are risky because AI behaviour changes according to context.
The product covers several areas: discovering unsanctioned “shadow AI” services and Model Context Protocol connections; analysing prompts, sessions and responses on supported platforms including Microsoft Copilot, Copilot Studio, ChatGPT Enterprise, Claude, Salesforce and AWS Bedrock; and monitoring agent identities, permissions, relationships and activity.
It also provides visibility into agent development in high-code and low-code environments, with the aim of identifying excessive permissions, misconfigurations and behavioural drift before deployment. Darktrace says its platform can block unsanctioned services or quarantine devices through integrations, while prompt analysis can flag attempted jailbreaks, sensitive-data exposure, potential indirect prompt injection and policy violations.
The announcement also highlights separate Darktrace research in which AI agents in a simulated corporate Active Directory environment independently moved from an impossible coding task to network scanning, credential discovery, lateral movement, credential dumping and modification of the benchmark. Darktrace says its products detected the activity, and an autonomous response initially blocked the agent’s connections.
In one test, SECURE AI gave the session an overall score of 99% and identified 23 usernames and passwords. These findings come from controlled testing, not confirmed attacks against customer environments. Darktrace says SECURE AI is generally available.