AI agent supply-chain security

Every agent
obeys someone.

Inspect the instructions, skills, and tools your AI agents trust. Find prompt injection, hard-coded secrets, and unsafe commands before your agents run them.

  • Read-only inspection
  • Secrets redacted in reports
  • No cloning or code execution

The platform in numbers

Inspection at a glance

Explore the fingerprint registry
Files inspected
-
All-time activity, including repeat scans
Unique skill snapshots
-
Registered HIGH/CRITICAL snapshots
Unique security findings
-
Distinct findings, not repeat scan hits
- detection rules 9 agent ecosystems How we evaluate

Evidence of inspection, not a safety rating. Snapshots can be partial; content versions are not unique people, publishers or approved skills.

Registry breakdown & counting methodology
Complete supported skill versions
-
MCP definition snapshots
-

Only snapshots with HIGH or CRITICAL findings are registered. New unique finding counts include only HIGH/CRITICAL findings; historical finding/evaluation counts are preserved and may include lower severities. Complete supported versions have all listed package files inspected at a pinned GitHub commit. MCP snapshots cover saved tool definitions, not running servers. Finding identities bind a content version, detector and evidence location. Minor content edits create new versions; historical unrecorded scans are not backfilled.

Loading registry details...

Scanner

Inspect before your agent obeys

Upload agent files, paste a skill, or point at a repository. Results appear in seconds.

Ready
Select a folder to keep paths like .claude/settings.json intact.

    Signed skill package? Upload the whole folder, including skill.oms.sig. NVIDIA-style OMS verification requires operator-configured publisher certificates. Pasted content alone is unsigned. GitHub directory URLs such as /tree/main/skills restrict the scan to that directory.

    Automatic high-risk fingerprint tracking and retention. Only snapshots with HIGH or CRITICAL findings publish fingerprints, self-declared frontmatter names, redacted finding excerpts and assessment summaries indefinitely, including pasted content, private repositories and uploads. Clean, LOW-only and MEDIUM-only snapshots are not centrally registered. Names and excerpts can disclose private code or identifiers; detected-secret redaction is not a guarantee of confidentiality. Source paths, repository URLs and full originals are not separate public fields. High-risk originals are retained privately for 30 days. Access tokens are never retained. Only submit material you can publish.

    Seen this skill before?

    Search the central registry for a content fingerprint and its latest stored assessment, including fingerprints registered by private scans.

    How it works

    Three steps, zero trust

    1. 01

      Collect

      Upload files, paste skill content, or name a GitHub, GitLab or Azure DevOps repository. For repositories, AgentLens lists the commit's file tree and inspects only files an agent would load.

    2. 02

      Inspect

      An isolated scanner applies injection, credential, execution, and exfiltration rules. Nothing is run, rendered, or followed.

    3. 03

      Report

      Every signal is mapped to a file, line or decoded JSON pointer, severity, MITRE ATLAS technique, and research reference. Detected secret values are redacted before they reach your screen.

    Explore

    Explore AgentLens

    Defensive inspection informed by SkillVenom red-team research and samples.