Darktrace Releases SECURE AI to Monitor Sessions and Agent Permissions

Companies using AI tools across their workforce now have a new monitoring option: Darktrace has made SECURE AI generally available as a standalone product or part of its Behavioral Defense Platform. The offering is built to identify risky AI activity, including sensitive-data exposure, jailbreak attempts, unsanctioned services and agent misconfigurations.

The release addresses a visibility problem. Darktrace says telemetry from about 8,200 deployments showed that more than 80% of monitored customer accounts used generative AI services in August. The average organization used five AI providers, leaving security teams to manage approved assistants, employee-selected tools and internal agents at once.

From AI conversations to device controls

SECURE AI applies Darktrace’s behavioral-detection approach to corporate AI use. The company’s security products establish a picture of normal activity within each customer environment; this product uses that approach to flag activity that departs from that baseline. Its general-availability release includes integrations spanning AWS, Anthropic, Microsoft and OpenAI services.

The controls Darktrace says it adds

• It scans sessions in ChatGPT Enterprise, Claude, Microsoft Copilot and Amazon Bedrock in real time for jailbreak attempts, sensitive-data exposure and indirect prompt injection.
• It ranks sessions by risk, detects unsanctioned AI services and can block those services or quarantine a device through integrations including Microsoft Entra Global Secure Access.
• It maps AI agents to their human users, records what each agent can access and checks development environments for excessive permissions and misconfigurations while agents are being built.

The product reaches into agent creation

The development-side coverage is a notable part of the pitch. SECURE AI monitors environments including Amazon Bedrock and Microsoft Copilot Studio, where companies build agents, for permissions that are too broad and configuration mistakes. Mapping an agent to a person and its allowed access is meant to give administrators a trail from automated activity back to an accountable user.

A policy layer beside the detection

Darktrace says customers can upload AI policies in plain language and have SECURE AI check them against regulatory and compliance frameworks. Darktrace did not identify the supported frameworks in its announcement.

A broader security partnership remains in development

The release also sits alongside Darktrace’s work with OpenAI through the Daybreak Defense Network. Darktrace is developing capabilities that combine OpenAI’s Daybreak models with its behavioral data. The stated goal is to show defenders which people and systems an AI-related incident has affected.

SECURE AI is available now to new and existing Darktrace customers, either on its own or through the Behavioral Defense Platform. For companies confronting a growing inventory of AI services and agents, the central question is less whether activity exists than how far monitoring and enforcement should extend into everyday use.