On 22 September 2026, U.K. cybersecurity firm Darktrace made SECURE AI generally available, bringing behavioral detection to enterprise use of AI tools and agents. The SiliconANGLE exclusive follows a February announcement and answers an operational fact: AI has spread through companies faster than security teams can inventory it.
Telemetry from about 8,200 Darktrace deployments shows more than 80% of monitored customer accounts used generative AI services in August, and the average organization dealt with five different AI providers that month. Early customer examples are blunt: one firm approved a single assistant only to find most employees using unauthorized services; another discovered nearly 90 agents built inside a low-code environment with no approval process; a third opened an immediate legal review after unmanaged multi-platform use by contractors.
SECURE AI applies Darktrace’s long-standing behavioral approach—learning what normal looks like for each customer—to AI traffic and sessions. General availability includes integrations with Amazon Web Services, Anthropic, Microsoft and OpenAI. Sessions in ChatGPT Enterprise, Claude, Microsoft Copilot and Amazon Bedrock are scanned in real time for jailbreak attempts, sensitive-data exposure and signs of indirect prompt injection, with risk ranking so analysts know where to start.
Shadow AI detection leans partly on SASE integrations such as Microsoft Entra Global Secure Access. Administrators can block unsanctioned services or quarantine a device. Companies can also upload AI policies in free-form text and check them against regulatory and compliance frameworks—a practical bridge to EU AI Act duties and internal NIST controls.
Beyond chatbots, the product maps AI agents to the human users behind them and records what each agent is permitted to reach. Coverage extends into development environments such as Amazon Bedrock and Microsoft Copilot Studio, catching excessive permissions and misconfigurations while an agent is still being built, not only after it is live.
CEO Ed Jennings, who took over in March, summarized the thesis: the security industry spent two decades cataloging known threats, but AI breaks that model. The next attack will not always be recognizable in advance; the signal that matters is when something starts behaving differently from what is normal for the organization. Anomaly over signature is the core of SECURE AI.
Darktrace is also a member of OpenAI’s Daybreak Defense Network. Under that program it is developing capabilities that pair OpenAI’s Daybreak models with its own behavioral data, aiming to show defenders which people and systems an AI-related incident has touched. McCall McIntyre, OpenAI’s head of global cyber partnerships, said working with partners such as Darktrace is critical to translating AI advances into practical defenses.
SECURE AI is on sale now to new and existing customers as a standalone product or as part of the Behavioral Defense Platform, and it is listed on AWS Marketplace and Microsoft Marketplace. For CISOs the commercial message is clear: blocking ChatGPT is not enough; you must see shadow agents, MCP permissions and behavioral drift.
September 2026 makes the launch especially timely. Frontier labs negotiate cross-lab safety testing, the European Commission is sending information requests to more than 30 providers, and parliaments are calling hearings on agentic incidents. Enterprises, meanwhile, are discovering that employees already deployed dozens of agents without governance.
SECURE AI does not by itself equal regulatory compliance, but it offers an operational sensor where blind spots used to sit. Inventorying agents, correlating them to users and cutting high-risk sessions is exactly the control auditors will demand as AI Act inspections intensify in HR, banking and healthcare.
Known limits remain: telemetry depends on connectors and SASE visibility; agents operating outside instrumented channels stay opaque; and a risk ranking does not replace human investigation. Still, moving from pilot to GA with integrations across the four major model providers marks an inflection in the AI-native security market.
For Neural Pulse, the most telling figure is not product marketing but telemetry: five AI providers per average organization and 80% generative adoption in monitored accounts. That confirms the agentic attack surface is already mainstream, not experimental.
In short, on 22 September 2026 Darktrace turned behavioral monitoring of enterprise AI into a GA product. Shadow AI, low-code agents and prompt injection leave the blog-post anecdote stage and become prioritized SOC alerts.
We will watch early public customer metrics, the real depth of Daybreak integrations, and whether other XDR/SASE vendors answer with equivalent suites before year-end.
