TechPrufer
Get started
Glean
Glean logoGlean · Feature trail

Security AI Agents

16 posts · Nov 2025 – Oct 2026

What this trail covers

Glean's security AI agents framework has evolved to address the challenges of traditional security architectures by providing a structured approach to evaluating and governing AI agents. This includes the AWARE framework, which offers shared decision criteria for CIOs and CISOs, focusing on intent, context, guardrails, runtime risk, and observability. The framework enables a repeatable process for assessing agent risk, moving from 'risk is too high for now' to approval by establishing clear technical controls for scanning, sandboxing, and credential brokering. This post details the implementation of these controls for skills, ensuring they are safe to use in the enterprise by scanning them before entry, using a credential broker at runtime, and implementing access controls to scale safely.

All Glean trails
  1. How Glean ensures skills stay secure and permissions awareThis post details Glean's approach to securing AI skills by implementing a multi-stage security model. It introduces a scanning process that evaluates skills for prompt injection, data exfiltration, and dangerous tool usage before they are deployed. At runtime, skills are executed within isolated sandboxes, and a credential broker mediates access to external systems, ensuring that skills only use permitted credentials and connect to approved destinations. Governance is enforced through access controls on sharing, administrator/moderator roles, and a draft-to-publish lifecycle with versioning and rollback. This architecture allows for the safe adoption of skills from external ecosystems like GitHub and MCP, ensuring enterprise control without sacrificing openness.
  2. AI Agents for Financial Compliance: How Enterprise AI Is Solving the Regulatory Reporting ChallengeThis post details the application of Glean's security AI agents framework, specifically the AWARE framework, to the domain of financial compliance. It explains how AI agents can be used for continuous regulatory monitoring, automated audit trail generation, and AML/KYC lifecycle management. The post emphasizes the importance of explainability, traceability, and human-in-the-loop controls for deploying AI agents in regulated environments, outlining a practical roadmap for organizations to get started with AI-driven compliance.
  3. Scale AI usage without runaway spend with Glean’s AI usage controlsIntroduces AI usage controls and Glean AI Gateway to manage AI costs and scale AI adoption responsibly. This includes features like usage limits (company-wide, user, department, agent level), alert-only and hard-cap enforcement, department and user-level usage visibility, and department-based and premium model governance. The AI Gateway acts as a control layer beneath AI tools, managing model access, token usage, attribution, and spend across various AI surfaces.
  4. 6 AI agents transforming finance workflowsThis post details six specific use cases for AI agents within finance workflows, demonstrating how they can be built on Glean's platform to address challenges like reconciliation, variance analysis, audit prep, policy compliance, and reporting. It emphasizes the agents' ability to understand intent, access verified sources, and deliver context-rich, permission-aware responses by integrating with enterprise systems like ERPs, contract repositories, and shared drives. The examples illustrate how these agents reduce manual effort, improve accuracy, and support compliance, aligning with the broader security AI agents framework.
  5. AI is its own best safeguard: Glean uses AI to detect jailbreak attempts at 97.8% accuracyIntroduced new AI security detection models that achieve 97.8% prompt injection detection, 93.5% toxic content detection, and 94.3% malicious code detection on leading benchmarks. Developed enterprise-specific fine-tuning for these models using publicly available datasets. Implemented a dual model approach to reduce false positives, achieving a 3.0% false positive rate on prompt injection attacks. Enhanced the AI security dashboard to provide context for violations, including surrounding context, prompt, source file, agent identifier, user information, and chat session.
  7. Secure generative AI for the enterprise requires the right permissions structureThis post details how Glean ensures enterprise-readiness for generative AI by prioritizing a robust permissioning structure. It explains the importance of permission awareness for AI assistants, adhering to the principle of least privilege, and handling complex, nuanced permissioning across diverse datasources. The post highlights challenges like API rate limits, data freshness, and the cost of indexing individual document permissions, advocating for a unified permissions model. It also discusses edge cases in datasource permissions (e.g., GDrive's public visibility after visit, temporary access, non-flattened criteria) and the engineering effort required to build a resilient system that accounts for these complexities at scale.
  8. What is AI security?This post details common vulnerabilities in AI systems such as data poisoning, lack of transparency, and prompt injection attacks (direct, indirect, and chained). It also outlines mitigation strategies during AI deployment, including securing the AI lifecycle (data integrity, monitoring for drift) and input validation/prompt handling. The adoption of a zero-trust model with continuous verification, granular access controls, and endpoint/network security is emphasized. Best practices and frameworks like NIST AI Risk Management Framework, OWASP Top 10 for LLM Applications, MITRE ATLAS, and ISO/IEC standards for AI security are highlighted.
  9. How to build agents that actually work: A practical guide to evaluating AIThis post details a practical guide to evaluating AI agents, focusing on creating evalsets with real-world examples and defining success metrics like message quality (completeness, personalization, tone, groundedness). It outlines a systematic approach to improving agents by refining instructions, structuring context, tuning parameters, and selecting appropriate LLMs. The post also discusses scaling evaluations through automation and LLM grading, and integrating user feedback. This contributes to the 'Security AI Agents' thread by providing a methodology for ensuring agent reliability and trustworthiness, which is crucial for security applications.
  10. Glean AI Gateway: One control plane beneath every AI front doorThis post introduces Glean AI Gateway, a new control plane that sits between AI surfaces and models/tools. It enables centralized governance of model/MCP access, token usage, and security across multiple AI front doors. Key technical contributions include cost control mechanisms (budget limits, quotas, model fallback), enhanced visibility into token/tool usage and performance metrics, consistent application of data permissions and granular access controls, and integration with Glean's enterprise graph for improved context. It also details initial LLM and MCP support.
  11. 7 essential guardrail decisions for deploying enterprise AI agents successfullyThis post outlines seven key guardrail decisions for deploying enterprise AI agents, providing a structured framework for secure, compliant, and financially sustainable deployments. It details considerations for model choice and hosting (managed APIs, private cloud, on-premises), decision architecture and RAG, tool permissions and RBAC, human-in-the-loop and risk tiers (low, medium, high), observability and audit logging, input/output validation with content safety, and cost controls and financial operations (FinOps). It emphasizes the importance of aligning AI agent capabilities with enterprise governance goals and data protection policies.
  12. How to Deploy AI Agents With Confidence: Introducing AWARE and New Glean Protect CapabilitiesThis post introduces the AWARE framework (Actor Intent, Work Context, Autonomous Guardrails, Real-Time Risk Scoring & Blocking, Ecosystem Observability) for governing AI agents. It details new Glean Protect capabilities: restricted topics policies and agent alignment models for 'Autonomous Guardrails', enhanced sensitive content policies and models for 'Work Context', and improved triage workflows and integrations (Tines, Cortex Cloud DSPM, Prisma AIRS) for 'Ecosystem Observability'. It also announces customer-managed deployment for single-tenant environments, offering control over code deployment, private connectivity, and customer-managed keys.
  13. 8 AI agents to help software engineering teams ship and fix fasterThis post introduces eight new AI agents for software engineering teams: Pull Request Review, Resolve GitHub PR Feedback, Resolve Jira Ticket, Spec to Implementation PR, Engineering Project Onboarding, Launch Documentation, Engineering Self-Evaluation, and Engineering Standup. These agents are designed to automate and streamline common, time-consuming engineering tasks by leveraging context from various development tools.
  14. 5 metrics to measure AI-generated answers' decision-making impactThis post details five metrics (accuracy, relevance, coherence, helpfulness, and user trust) for measuring the impact of AI-generated answers on decision-making. It provides technical details on how to measure these metrics, including specific scores like faithfulness, BLEU, ROUGE, perplexity, and readability, as well as user feedback mechanisms and structured scoring rubrics. The post also discusses the importance of retrieval augmented generation for grounding answers in company data and outlines strategies for continuous monitoring and improvement.
  15. 11 AI agents powering every stage of security operationsThis post details 11 specific AI agents that power different stages of the security operations alert lifecycle. It elaborates on the 'Device posture check' agent, which connects to endpoint and configuration management tools via Glean's MCP host to pull configuration and vulnerability data, correlating it with EDRs and management systems to flag noncompliant assets and remediation priorities. It also describes the 'Identify incident experts' agent, which uses Glean's enterprise graph to identify subject-matter experts based on historical post-mortems, tickets, and documentation. The 'Incident correlation' agent links alerts and tickets from systems like Jira, ServiceNow, and Slack to uncover root causes. The 'Open security-related tickets' agent pulls open tickets from Jira and ServiceNow, grouping them by asset, severity, or customer. The 'Security issue impact' agent analyzes telemetry, asset inventories, and historical ticket data to determine blast radius and business impact. The 'Incident response process' agent guides teams through approved playbooks or drafts tailored templates. The 'Incident response documentation' agent automates the creation of incident summaries, evidence logs, and post-incident reports by recording activity timelines and collaboration threads. Finally, the 'Post-mortem creation and containment' agent standardizes incident recovery by automating documentation and guiding containment steps.