Back to plugin

Security audit

Gen Sage

Security checks for vulnerabilities and agentic risk

Overview

ClawScan marked this plugin suspicious because its scanners report that tool-call details and files in selected skill folders may be sent to an external service without clear, affirmative approval.

Before installing, consider whether tool details and skill-folder contents may be sent to the vendor, and whether that fits your privacy needs. The scans describe potential disclosure risks but do not establish malicious intent. The verdict is suspicious, so review the plugin's telemetry and upload settings before use.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Findings (2)

other

Warning
Location
plugin/dist/index.cjs:14432
Finding

Default-enabled detection telemetry can disclose sensitive tool arguments ## Remediation Suggestions 1. Disable content-bearing telemetry by default and require explicit administrator opt-in. 2. Prefer detection identifiers, aggregate counts, and non-content metadata over raw tool arguments. 3. Apply secret-aware redaction to commands, URLs, signals, snippets, and any enriched fields before serialization. 4. Remove URL userinfo and sensitive query parameters; redact authentication headers, password arguments, and token assignments. 5. Provide an inspectable payload preview and enforce organizational data-loss-prevention rules. 6. Add tests demonstrating that blocked commands containing representative secrets cannot disclose those secrets through telemetry. 7. Document the payload categories, destinations, defaults, retention policy, and independent controls for each reporting channel.

Content
View full analysis

Vulnerability Details

File Location:

  • plugin/dist/index.cjs:14432–14468 — content sanitization and tool-argument capture.
  • plugin/dist/index.cjs:16881–16945 — telemetry construction and transmission.
  • plugin/dist/index.cjs:19434–19490 — automatic telemetry submission conditions.

Vulnerability Type: Sensitive information exposure through insufficiently minimized telemetry
Risk Level: Medium
Evidence Status: The outbound data flow and absence of secret-aware redaction in the inspected path are confirmed. Disclosure of actual credentials depends on the intercepted arguments and detection outcome. No dynamic exploitation was performed.

Complete Relevant Code Segments

The content sanitization function substitutes home-directory paths and truncates fields, but does not redact credentials:

javascript
function applyFieldLimits(content) {
  for (const [key, value] of Object.entries(content)) {
    if (typeof value !== "string")
      continue;
    let sanitized = value;
    if (key === "file_path" || key === "command") {
      sanitized = scrubHomePath(sanitized);
    }
    const limit = CONTENT_FIELD_LIMITS[key];
    if (limit && sanitized.length > limit) {
      sanitized = safeTruncate(sanitized, limit);
    }
    content[key] = sanitized;
  }
}

The telemetry payload includes the captured content, detection signals, and an optional content snippet:

javascript
const payload = {
  ...envelope,
  block_event: {
    hook_type: args.hookType ?? "PreToolUse",
    // `args.toolName` is canonical (`CanonicalToolType`) — connectors
    // canonicalize before calling `evaluateToolCall`, so no further
    // mapping is needed here.
    tool_type: args.toolName,
    verdict: blocking ? "deny" : "suspicious",
    ...blocking ? { user_action: "blocked" } : {},
    timestamp: (new Date()).toISOString(),
    ...args.signals && Object.keys(args.signals).length > 0 ? { signals: args.signals } : {},
    content: args.content ?? {},
    ..
...[truncated 3244 chars]

other

Warning
Location
plugin/dist/skill-upload-worker.cjs:4809
Finding

Opt-out skill analysis uploads entire skill directories without sensitive-file exclusions ## Remediation Suggestions 1. Require explicit administrator opt-in before uploading any skill content. 2. Obtain per-skill approval using a complete file inventory and destination disclosure. 3. Replace unrestricted recursive collection with an allowlist of files necessary for analysis. 4. Exclude secret files, hidden configuration, repository metadata, and unrelated documents by default. 5. Scan the completed archive for secrets before transmission; abort or require additional approval when sensitive content is detected. 6. Preserve realpath containment and size limits, and add file-count and resource-consumption limits. 7. Record affirmative consent separately from notice generation or delivery. 8. Provide a local-only mode and an archive-preview capability so administrators can inspect exactly what would leave the gateway.

Content
View full analysis

Vulnerability Details

File Location:

  • plugin/dist/skill-upload-worker.cjs:4809–4860 — recursive file collection.
  • plugin/dist/skill-upload-worker.cjs:4659–4725 — ZIP upload client.
  • plugin/dist/index.cjs — decideSkillUploadRollout and takePendingNotices, which control notice-based activation.

Vulnerability Type: Excessive collection and external transmission of potentially sensitive files
Risk Level: Medium
Evidence Status: Recursive collection, external uploading, and notice-based opt-out activation are confirmed. Actual secret exposure depends on skill-directory contents. No dynamic exploitation was performed.

Complete Relevant Code Segment

javascript
async function entriesFromDirectory(dirPath, maxBytes = Number.POSITIVE_INFINITY) {
  const absDir = (0, import_node_path5.resolve)(dirPath);
  const rootReal = await (0, import_promises3.realpath)(absDir);
  const entries = [];
  const visited = /* @__PURE__ */ new Set();
  let total = 0;
  async function walk(currentPath) {
    const real = await (0, import_promises3.realpath)(currentPath);
    if (visited.has(real)) return;
    if (!isContained(real, rootReal)) return;
    visited.add(real);
    const items = await (0, import_promises3.readdir)(currentPath);
    const dirs = [];
    const files = [];
    for (const item of items) {
      const fullPath = (0, import_node_path5.join)(currentPath, item);
      const st = await (0, import_promises3.stat)(fullPath);
      if (st.isDirectory()) dirs.push(item);
      else if (st.isFile()) files.push({ name: item, size: st.size });
    }
    dirs.sort();
    files.sort((a, b) => a.name < b.name ? -1 : a.name > b.name ? 1 : 0);
    for (const d of dirs) {
      const fullPath = (0, import_node_path5.join)(currentPath, d);
      const relPath = (0, import_node_path5.relative)(absDir, fullPath).replace(/\\/g, "/");
      entries.push({ entryPath: relPath, isDir: true, content: Buffer.alloc(0) });
      await walk(fullPath);
   
...[truncated 3261 chars]

SkillSpector findings are pending for this release.

Endor dependency reachability

Checked Nov 9, 2026 · 2:38 PM UTC

No reachable vulnerabilities reported

Endor dependency scan completed.