Prompt injection
Instruction overrides, fake system tags, hidden comments, and invisible Unicode that redirect agent behavior.
Detection coverage
Inspect instructions, configuration, and the permissions between them.
Instruction overrides, fake system tags, hidden comments, and invisible Unicode that redirect agent behavior.
Cloud, AI-provider, GitHub, and npm tokens, private keys, and credentialed URLs, redacted in every report.
Pipe-to-shell installs, risky hooks, MCP servers, auto-approval, and workspace tasks that run on open.
Instructions that ship files, environment variables, or credentials to external endpoints, DNS, or images.
Understands the files these agents load
Evaluation transparency
AgentLens reports static signals, not a malware verdict or safety certification. Original submissions are not published in a directory. Sanitized assessments are publicly available through exact snapshot-fingerprint or indexed file-text-hash lookup.
The signal risk score runs from 0 to 100. Higher means more distinct rule evidence, not a probability of malware. Each rule contributes only its highest observed severity: CRITICAL 40, HIGH 20, MEDIUM 5, LOW 1. The sum is capped at 100. Repeated matches of one rule do not inflate the score.
Partial scans and scans with no target files are not scored. A complete scan only covers the selected, supported files, not all behavior of a skill or repository. A zero score means no current rules matched, not that the skill is approved.
Compare scores only with the same methodology version, detector fingerprint and scan scope. A changed rule set can change a score without any change to a skill.
Static pattern analysis and embedded-secret pattern detection, plus bounded recursive base64, UTF-16LE base64, hex and percent decoding. Saved JSON MCP tool definitions are inspected as decoded data with JSON-pointer evidence. Schema and sensitive-input observations are review notes, not scored vulnerabilities. No MCP server is contacted.
NVIDIA-style skill.oms.sig bundles can verify submitted package bytes and publisher certificate identity against operator-configured trust anchors. Strict checks reject tampering and unsigned additions; missing or unsupported files prevent verification. Reputation, online revocation and timestamp verification are not assessed. A valid signature is not a safety certification.
No sandbox execution, live credential checks, publisher reputation checks, remote-content inspection or installed dependency analysis. AgentLens does not execute submitted files.
Reports include the repository commit when available, SHA256 hashes of inspected UTF-8 text, an input-manifest hash and a detector fingerprint. An optional UTF-8 BOM is removed before hashing. These are not signed attestations.
Encoded-text inspection follows up to three layers with shared per-file limits of 128 candidates, 32 payloads and 64 KiB of decoded text. Reaching a limit makes the scan partial and unscored. Findings identify the containing source line or JSON pointer, transformation chain and decoded line; decoded line numbers are not presented as physical source lines.
Export SARIF 2.1.0 from the report for code-scanning tools. The browser builds the export locally, excluding snippets and original content. It retains evidence locations, fingerprints and incomplete-scan status, without inventing CVSS scores. Review paths, pointers and other metadata before publishing.
A skill snapshot groups inspected files under a SKILL.md folder. Its content ID binds package-relative paths, raw-byte hashes and file modes where known. An evaluation ID also binds the detector, options and completeness; finding IDs bind evidence locations within that evaluation. Matching copies with equivalent metadata count once, while content edits create new versions. Unknown file modes are not treated as known modes.
Only eligible pinned GitHub packages with every listed file inspected can count as complete supported versions. Uploads and other repository providers remain snapshots. Missing files, unsupported assets, nested skills and partial scans prevent complete-parent claims. Completeness alone does not establish a valid signature or safe behavior, and is not exhaustive dependency resolution.
Only snapshots with HIGH or CRITICAL findings register identities and assessment summaries indefinitely, including private repositories, pasted content and uploads. Clean, LOW-only and MEDIUM-only snapshots remain in the scan report but are not centrally stored. Central fingerprint search retrieves stored metadata and redacted finding excerpts, never full files or repository contents. Detected secrets are masked, but names or excerpts may still disclose private code or identifiers. Full originals remain private. High-risk originals are retained for 30 days. Access tokens are never retained. New unique finding counts include only HIGH/CRITICAL findings; historical finding/evaluation counters and digests are preserved and may include lower severities. Startup removes non-high/critical and unclassifiable legacy snapshot records and their lookup links. Public counts are cached shard snapshots, not unique people, publishers or adoption figures. Storage and excerpt counts are bounded; errors and omissions are disclosed.
Read the machine-readable methodology and rule categories. Loading detector identity...