Agent security
Use agent security controls to assess risks in agent configuration and govern activity during agent runs. Choose the control that matches the risk you need to address.
Configuration risks
Agent scan evaluates an agent definition against scan policies. It checks descriptions, instructions, tools, data access, and approval settings for risks such as excessive data access or write tools that do not match the stated purpose.
Create scan policies, review findings, and require publication approval for agents that match specific rules. See Agent scan.
Runtime issues
AI security policies evaluate user prompts, retrieved content, and model responses during agent runs. They detect risks such as prompt injection, malicious code, and harmful content. The agent alignment check also monitors whether write tools match the user request and agent instructions.
Choose which agent types each policy covers and how Glean responds to violations. See Configuring Policies.
Use the Findings Dashboard to investigate individual violations. Review trends and recurring risks in AI Security Insights.
Agent access policies
Agent access policies evaluate tool arguments before a tool runs and tool output after it returns. Use them to block unsafe tool calls, filter sensitive items from results, or flag activity for review.
Scope policies to users, groups, agents, and tools. For example, block posts to public Slack channels or filter confidential documents from search results. See About agent access policies.
These controls complement existing source permissions and agent access settings. Agent scans assess configuration risk, while runtime policies evaluate activity during a run.
See also
- Policies and examples: Review the risks that runtime AI security policies detect
- Manage agent access: Configure agent roles, sharing, and publishing permissions