Darktrace ships SECURE AI to find the AI services its customers never approved

Somebody at one company wrote a policy naming a single approved AI assistant, and most of the staff used unauthorized services anyway, according to Darktrace's account of an early deployment ^[5].

The scale case is built on Darktrace's own sensors. More than 80% of about 8,200 monitored deployments showed generative AI use in August, which works out to upwards of 6,560 estates on those numbers ^[2][22]. The average organization dealt with five different AI providers in the same month ^[3]. Both figures count use. The unapproved share is carried by three examples from early customer deployments ^[6][5][4].

Five providers on average, four platforms with session-level inspection ^[8][3]. At least one provider in the average estate therefore sits outside real-time inspection and falls to the shadow AI detection path, which runs partly through secure access service edge integrations such as Microsoft Entra Global Secure Access ^[23][10].

The agent work is the part a manual inventory will not reproduce. Darktrace maps each agent to the human user behind it and records what that agent is permitted to reach, including inside Amazon Bedrock and Microsoft Copilot Studio while the agent is still being built ^[12][13]. The early customer with nearly 90 agents in one low-code environment, all created without any approval process, has nearly 90 sets of permissions and no owner of record ^[4].

Ed Jennings, who became Darktrace's president and chief executive in March, said the industry spent two decades cataloging known threats. "AI breaks that model," he said ^[15][16]. The signal he points to instead is behavioral: something that "starts behaving differently from what is normal for the organization" ^[17]. That approach has a standing staffing cost, and the risk ranking on each session is the product's own acknowledgment of it, since nobody reviews every session ^[9].

Buyers should also separate what ships from what is described. Darktrace is developing capabilities that pair OpenAI's Daybreak models with its own behavioral data to show defenders which people and systems an AI-related incident has touched ^[18]. Policy checking in the shipping product means uploading your own AI policies as free-form text and having the software compare them against regulatory and compliance frameworks; enforcement stays with blocking a service or quarantining a device ^[11][10]. SiliconANGLE did not report a price, standalone or as part of the platform ^[21].

The count is available before the budget conversation. Thirty days of proxy or SASE logs, a list of every AI provider with traffic in them, and an export of the agent list from each low-code tenant. Then two questions per item: can the sessions be inspected, and does a named person own it. Anything failing both is what the money is for, and in the average estate Darktrace describes, there is at least one such provider ^[23].