Tool access and scope
Review which tools the agent can call and whether their permissions match the actions it is meant to perform.
Galen SecurityKnow what your agent
GALEN SECURITY / AI AGENT SECURITY AUDIT
Know what your agent
can actually reach.
Review the tools, APIs, permissions, and data behind your agent. Galen’s web application analysis examines the supported code paths behind those connections, giving your team evidence of which actions are reachable and which controls enforce the intended boundaries.
Available now: Galen Policy Beta and Managed Galen Security Audits. Self-service Galen Security is coming soon.
Inspectable evidenceRemediation and validation
WHAT THE AUDIT COVERS
Every connected tool opens a path.
An agent’s risk depends on the authority behind its tool calls. Review what each connection allows, whose identity it uses, and which controls limit the result.
MCP and function-calling paths
Examine the endpoints, authentication, and permission boundaries behind MCP servers and callable functions.
Data and tenant boundaries
Trace access to customer records, files, repositories, and connected systems for unintended exposure.
Prompt-injection impact
Assess the data or actions that could become reachable if untrusted input changes the agent’s behavior.
Sensitive actions and approvals
Review whether exports, account changes, administrative operations, and other high-impact actions require the intended authority.
Remediation and re-evaluation
Use the retained tool path, permission evidence, and approved access rules to guide corrections and later checks. Changes to the agent or coding assistant do not replace that security reference.
A FINDING IN CONTEXT
A helpful agent with an overpowered tool.
A support agent can call an export tool. The tool uses a shared credential and does not enforce the requesting customer’s tenant.
Illustrative scenario, not a live scan or actual customer finding.
FOLLOW THE PATH
- 01Support request
- 02Shared export tool
- 03Cross-customer records
What makes it risky
The tool grants access beyond the customer boundary the agent is expected to respect.
What to change and recheck
Enforce the caller’s tenant and permissions in the underlying tool or API, then recheck the agent-to-data path.
FROM FINDING TO CORRECTION
Give your team
something it can act on.
Review the evidence, make a focused correction, and check the result against the original risk.
See how Galen Security works- Findings linked to the affected tool, API, or agent workflow.
- Evidence of the relevant access path and permission boundary.
- Remediation direction for scopes, controls, and sensitive actions.
- Validation status after the updated workflow is reviewed.
A FEW PRACTICAL QUESTIONS
Plan the review.
Is this only a prompt-injection review?
No. The review focuses on the tools, data, actions, and permission boundaries an agent can reach. Source-backed analysis examines supported implementation paths; prompt-injection behavior and deployed permissions may need additional scoped testing or context. An instruction asking a model to respect a boundary is not proof that the underlying tool enforces it.
Can MCP servers be included?
Yes. MCP servers and function-calling endpoints can be included in the agreed review scope, along with the underlying APIs and permissions.
Can Telhawk lead the engagement?
Yes. A Managed Security Audit can define the agent workflows, connected systems, access requirements, remediation review, and reporting for a Telhawk-led engagement.
RELATED AUDITS
Follow the risk
across your application.
MANAGED NOW / SELF-SERVICE COMING SOON
Choose who leads
the review.
Telhawk-led managed engagements are available now. Self-service Galen Security is coming soon.