Agent vulnerabilities span several layers
A tool-using agent can inherit conventional dependency vulnerabilities, MCP transport weaknesses, unsafe tool descriptions, excessive permissions, exposed secrets, and dangerous combinations of otherwise legitimate actions. No single package lookup establishes the whole risk picture.
A defensible assessment starts with inventory and evidence. It then separates known-vulnerable component presence, reachable behavior, permission design, policy coverage, and runtime containment into distinct findings.
- Inventory the agent client, MCP servers, tools, transports, and dependencies.
- Map credentials and privileges to the tools that can use them.
- Review public advisories and validate runtime relevance separately.
- Record limitations, test boundaries, and remediation ownership.
What a Bindfort assessment provides
The current Bindfort assessment path concentrates on installed-tree vulnerability evidence and the policy-and-receipt boundary for a selected MCP flow. It is designed for guided review with non-sensitive inputs.
Automated secret discovery, broad configuration analysis, continuous attack-path analysis, and production containment are future capabilities. Public pages avoid presenting those items as shipped features.
Product status
Verified today and clearly separated from roadmap
- Installed dependency-tree scanning for selected MCP server inputs.
- Evidence summaries with advisory and resolved-version context.
- Controlled policy tests and receipt integrity verification.
- Explicit limitations and separation of presence from reachability.
- Broader agent configuration and secret-exposure analysis.
- Automated attack-path and cross-tool sequence analysis.
- Continuous scanning and runtime response integrations.
Frequently asked questions
Clear answers for evaluation
What does an AI agent vulnerability scanner check?
It should check the agent’s tools, MCP servers, installed dependencies, credentials, permissions, configuration, external content paths, and runtime controls.
Is Bindfort a complete automated agent scanner today?
No. The verified offering is a guided MCP dependency-tree and policy-evidence evaluation. Broader automated analysis remains roadmap work.
Will a scan expose source code or secrets?
The public website never requests source code or secrets. Any private assessment must use an explicitly agreed scope and should begin with the minimum metadata needed for the review.