Onyx and Glean integrate controls for enterprise AI agents
Onyx and Glean's beta integration inventories enterprise AI agents and checks prompts, responses and proposed tool executions against policy before actions reach business systems.
Why it matters
- Glean says its connector ecosystem spans more than 275 applications and data sources.
- The Onyx Security integration is listed as beta in Glean's connector catalog.
- Discovery reads agent inventory and security posture through the Glean API without executing agent actions.
- Runtime enforcement uses Glean hooks to evaluate prompts, responses and proposed tool executions against enterprise policies.
- Onyx offers gateway capabilities for AI and Model Context Protocol traffic and a Guardian Agent for risk detection and remediation.
By Elena Vasquez
Glean says its connector ecosystem spans more than 275 applications and data sources. Its new Onyx Security integration is listed as beta.
Onyx Security and Glean announced a partnership to secure and govern enterprise AI agents as they gain access to company information and business tools. The integration joins Glean's enterprise AI platform with Onyx's secure AI control plane. It brings agent discovery, risk assessment, policy enforcement and runtime protection into one governance approach.
The companies frame the enterprise need around a sharp distinction. An agent may have permission to retrieve information, yet still need additional controls over what it does with that information. A workflow can move from gathering context to sending a message, updating a record or invoking another tool.
For both companies, that distinction makes runtime oversight a practical enterprise control. It gives security teams a way to govern actions as agents become more capable. For enterprises, the policy stakes sit in the authority those agents receive. Approved access can lead to external disclosures, changed records and tool calls with business consequences.
“As AI agents take on more complex work across the enterprise, they need access to the right context, with the right controls in place,” said Zubin Irani, VP of Partnerships at Glean.
What does the integration actually control?
According to Onyx's technical description, the partnership has two enforcement paths. Discovery and runtime controls do not run through the same connection.
- Discovery: Onyx says the Glean API reads agent inventory and security posture. Security teams can review agents and their access, but the discovery connection cannot execute agent actions.
- Runtime enforcement: Glean's hooks provide supported checkpoints where Onyx can evaluate conversation content and proposed tool executions against enterprise policies.
Onyx says the controls assess prompts, responses and agent actions. Their stated aim is to block unsafe requests and steer unintended behavior back toward the authorized task. The check happens before a supported tool action reaches a business system.
Onyx illustrates the difference with an enterprise workflow. Access to internal account information may be appropriate while an agent prepares a customer meeting. Sending confidential pricing information outside the company may violate policy. An action checkpoint can assess that proposed disclosure before the tool executes it.
The partnership documentation also identifies instructions embedded in external content. A document or tool response can attempt to redirect an otherwise legitimate workflow. Runtime checks are intended to govern that action after retrieval, not merely the initial request for data.
Where does Glean's enterprise context come from?
Glean provides the platform for building, deploying and operating enterprise agents. Its Enterprise Context layer connects business applications and indexes company information. Its Enterprise Graph maps relationships among people, content and activity.
Glean says the connector ecosystem includes documents, tickets, chat and code. The company says this context can link a customer account with its support history and the people responsible for that relationship. Permission-aware retrieval then determines which information the user can access as the agent works.
Glean also supplies a security foundation through Glean Protect. Glean describes its coverage this way:
- User and agent identity.
- Source-system permissions and sensitive content.
- Runtime threats and action alignment.
- Audit events that can feed security monitoring systems.
The Onyx integration adds a separate enterprise-wide governance and risk layer around Glean agents and other supported AI environments. In the partnership, Glean supplies context and checkpoints; Onyx supplies the control framework.
Why must governance continue after approval?
Onyx says the integration lets security teams discover Glean agents, inspect their access and capabilities, establish ownership and business purpose, and assign risk classifications. Those records can inform enterprise AI governance and compliance policies.
Onyx's partnership description says ongoing assessment covers changes to permissions, tools and data access. An agent approved for one workflow may later gain a connector or capability. That can change its risk profile even when its name and owner stay the same.
The partnership provides centralized audit evidence for investigating activity and reviewing how policies were applied. It treats governance as continuous, rather than attaching a one-time approval to an agent's name.
“As agents gain greater access and autonomy, security teams need visibility and control over what they can do,” said Gil Elbaz, Co-founder and Chief AI Officer at Onyx Security. “By combining Glean’s enterprise context with Onyx’s security control plane, customers can govern agent activity consistently and confidently across their AI environment.”
What sits behind Onyx's control plane?
Onyx positions its secure AI control plane across SaaS applications, cloud services, endpoints and custom code. The company says the platform combines agent observability, security posture management, governance and runtime safeguards. It also describes session replay and audit trails for investigations.
Onyx lists natural-language policy controls and integrations with security information and event management and security orchestration systems. The company says these tools connect agent oversight with the wider security stack.
Onyx also offers gateway capabilities for AI and Model Context Protocol traffic. A supervisory Guardian Agent is designed to identify risks and remediate issues across the platform. Onyx says these capabilities bring Glean activity into a control framework that security teams can use across multiple AI systems.
What does beta status mean for deployment?
Glean's connector catalog lists the integration with a beta label. Glean's listing names four scope areas:
- Agent inventory.
- Posture management.
- Runtime visibility.
- Controls on prompts, model responses and tool execution.
According to Onyx, runtime enforcement depends on supported checkpoints. Onyx and Glean identify workflow coverage as a deployment consideration. Teams need to know whether a relevant prompt, response or proposed tool execution passes through a checkpoint Onyx can inspect.
For enterprises expanding agent use across departments, the companies say the partnership's practical value depends on three things:
- Keeping the agent inventory current.
- Carrying ownership and risk decisions into runtime policies.
- Applying those policies where agents can affect business systems.
Without those links, inventory can show what an agent is meant to do without consistently governing what it actually does. The immediate test for the beta is operational: which workflows have supported checkpoints, and whether ownership and risk policies reach each runtime decision. That mapping will determine whether the integration can govern actions across business systems.
Source: unite.ai
More from Elena Vasquez
Show full bio
Market editor covering technology and product news in artificial intelligence and machine learning at AI In Context.
708 articles
Related articles
- Obsidian Security Launches Runtime Policy Engine for AI Agent Actions
- BlackFog survey: 47% report sensitive data exposure via AI
- ServiceNow Makes OpenAI a Preferred Intelligence Layer for 80 Billion Workflows
- Commvault and Zscaler Team Up to Govern and Recover Enterprise AI Agents
- 87% of Enterprise Developers Use Low-Code. That Is the Agent Security Problem