Model capability and safety evaluations

Glean gives your organization flexibility to choose the models that best fit its needs. The model hub provides access to more than 40 models from leading commercial and open-model providers. This page covers the controls admins have over those models, how Glean hosts them, and where to find independent evaluations of model capability and safety.

Control model availability

Admins control which models are available in their organization. The model hub supports both targeted and broad controls:

• Turn off individual models — remove specific models that you don't want users to access. See Exclude or restrict models.
• Turn off open models — on Glean Universal Model Key deployments, turn off open models by the region where the model creator is based: the United States, Europe, or China. Each region has one switch that also covers open models Glean adds from that region later. To opt out of all open models, turn off every region. See Manage open models by region.

These controls let your organization choose the set of models that fits its requirements for capability, governance, security, and risk tolerance.

Understand model hosting and data retention

Glean works with leading inference providers to give you access to the models in the model hub. For the providers Glean may use, see Glean subprocessors.

Glean's agreements with these providers prohibit them from training on Customer Data (defined in the Glean Terms of Service). Glean delivers requests under Zero Data Retention (ZDR), so providers do not log or retain Customer Data.

The exception is models with Limited Data Retention. Currently, Claude Fable is the only model in this category. Glean turns on a Limited Data Retention model only when your organization explicitly agrees to Glean's Limited Data Retention Addendum in the Admin console.

Open models developed by companies in China are served only from infrastructure located in the United States. Glean doesn't send customer data or requests to the companies that developed these models.

Compare model capability

For independent comparisons of model capability, see these resources:

• Artificial Analysis — independent analysis and comparisons of model intelligence, agentic performance, coding, cost, speed, and other capabilities
• Arena Agent Leaderboard — a live leaderboard based on real-world agent sessions and signals such as task completion, tool reliability, steerability, and tool hallucination

Compare model safety

For independent evaluations of model safety and alignment, see these resources:

• Cisco LLM Security Rankings — evaluates how models resist adversarial attacks, including direct malicious prompts, jailbreaks, prompt injection, and multi-turn manipulation
• Arena Agent Alignment Index — measures agent alignment signals such as unauthorized actions, false attribution, and deceptive completion

Interpret third-party evaluations

Third-party evaluations use different datasets, methodologies, model versions, providers, and measurement criteria. Results can change over time and may not predict performance or risk in every enterprise workflow.

Review the methodology and scope of each evaluation when you compare models. Glean provides these resources for reference. They don't replace your organization's own testing, governance review, or risk assessment.

Add a security layer with Glean Protect

Glean Protect provides an additional security layer across your Glean deployment. It includes foundational controls such as strict permissions enforcement, isolated cloud architecture, encryption, role-based access controls, and audit logs.

Glean Protect+ adds AI security guardrails that help detect and block prompt injection, jailbreak attempts, malicious code, and toxic content across Glean Assistant and Glean Agents. It also includes agent alignment models that check tools before they run. Use Protect and Protect+ alongside model selection, independent evaluations, and your organization's governance processes for defense in depth.

See About Glean Protect and Protect+.

See also

• Supported LLMs — the full list of supported models and their availability by hosting environment
• Configure LLMs in the Model Hub — choose which models are available and which models power each workflow