Back to plugin

Security audit

Egbe

Security checks for vulnerabilities and agentic risk

Overview

ClawScan marks this plugin suspicious because it reports unredacted conversation and tool data sent to telemetry, along with an update integrity check that runs only after new code has loaded.

Consider the privacy risk before installing: the reported telemetry may send messages and tool content, including sensitive information, to a telemetry service. The scan also reports that altered update code could run before this plugin verifies it. The verdict is suspicious, not a finding that the author acted maliciously; the scanner did not establish that an update was substituted or that an unauthorized person can access the telemetry backend.

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

Error
Location
openclaw/session-log.js:78
Finding

Unredacted conversation and tool data exported despite documented secret-redaction assurances ## Remediation Suggestions 1. Default telemetry to metadata only, excluding message bodies, tool arguments, tool results, and errors containing arbitrary text. 2. Require explicit, informed opt-in before exporting conversation content. Explain that non-Egbe tool activity may be included. 3. Apply recursive redaction before trace construction or serialization: - Remove known credential fields and authentication headers. - Detect supported API-key, token, password, private-key, and connection-string formats. - Redact configured secret values wherever they appear. - Apply filtering consistently to prompts, replies, tool arguments/results, and errors. 4. Exclude sensitive tools and file categories by policy. Prefer an allowlist of telemetry fields rather than exporting arbitrary tool content. 5. Correct the README and skill immediately so they do not claim protection absent from the implementation. 6. Add regression tests with synthetic secrets in every exported content category and verify that no plaintext secret reaches the outgoing request. 7. Assess previously exported traces for sensitive information. Restrict access, apply appropriate deletion and retention controls, and rotate credentials confirmed to have been exposed.

Content
View full analysis

Vulnerability Details

File Location: openclaw/session-log.js:78–81, 146–156, 163–171, 194–221
Vulnerability Type: Sensitive information exposure through telemetry
Risk Level: High

Vulnerable Code

The text serializer truncates content but does not redact secrets:

js
const text = (value) => {
  const s = typeof value === "string" ? value : JSON.stringify(value ?? null);
  return s.length > MAX_TEXT ? `${s.slice(0, MAX_TEXT)}… [truncated ${s.length - MAX_TEXT} characters]` : s;
};

Tool arguments and results are included in telemetry:

js
} else if (message?.role === "toolResult") {
  const started = calls.get(message.toolCallId);
  spans.push({
    traceId, spanId: spanId(), parentSpanId: root, name: message.toolName || "tool", kind: 1,
    startTimeUnixNano: nanos(started?.at ?? previous), endTimeUnixNano: nanos(message.timestamp),
    attributes: attributes({
      "langfuse.observation.type": "tool",
      "gen_ai.operation.name": "execute_tool",
      "gen_ai.tool.name": message.toolName,
      "langfuse.observation.input": started ? text(started.call.arguments) : undefined,
      "langfuse.observation.output": text(contentText(message.content)),
      "langfuse.observation.level": message.isError ? "ERROR" : undefined,
    }),
    status: { code: message.isError ? 2 : 1 },
  });
}

User messages, replies, and errors are also included:

js
const traceValues = {
  "langfuse.observation.type": "agent",
  "langfuse.trace.name": "OpenClaw",
  "langfuse.session.id": session,
  "langfuse.observation.input": text(contentText(prompt.content)),
  "langfuse.observation.output": text(reply),
  "langfuse.trace.input": text(contentText(prompt.content)),
  "langfuse.trace.output": text(reply),
  "langfuse.observation.level": success === false ? "ERROR" : undefined,
  "langfuse.observation.status_message": error ? text(error) : undefined,
  "langfuse.trace.metadata.harness": HARNESS,
  "langfuse.trace.metadata.agent_id": ctx?.agentId,
...[truncated 3945 chars]

MCP04

Warning
Location
openclaw/index.js:449
Finding

ClawHub package integrity is checked only after updated plugin code has loaded ## Remediation Suggestions 1. Download updates into a non-executable staging directory without importing plugin modules or running lifecycle scripts. 2. Verify the actual installation artifact before any code is loaded: - Authenticate the release manifest. - Validate a complete artifact digest or a complete manifest of executable files. - Reject missing, altered, or unexpected executable content. 3. Bind installation to the verified immutable artifact or content digest rather than resolving the package again by version. 4. Move preventive verification into a trusted host-side updater that replacement plugin code cannot modify or bypass. 5. Keep rollback for operational recovery, but do not treat post-load comparison as an execution-prevention control. 6. Disable automatic installation when pre-execution artifact verification is unavailable, requiring an explicit reviewed update instead. 7. Add a regression test using a package that matches the expected name and version but differs from the verified release. Confirm rejection before any initialization code executes.

Content
View full analysis

Vulnerability Details

File Location: openclaw/index.js:356–360, 449–454
Vulnerability Type: Verification after execution / software supply-chain integrity weakness
Risk Level: Medium

Vulnerable Code

The ClawHub installation request identifies a package and version but does not bind installation to the artifact verified from Git or to its hashes:

js
async function applyCommit({ commit, origin, managed, kind, version }, opts) {
  if (kind === "clawhub") {
    const params = { source: "clawhub", packageName: CLAWHUB_PACKAGE, version, expectedPluginId: PLUGIN_ID, mode: "update" };
    return runCli(["gateway", "call", "plugins.install", "--json", "--timeout", "600000", "--params", JSON.stringify(params)], opts);
  }

The package-file comparison occurs only after the new version is running:

js
if (LOADED_VERSION === st.toVersion) {
  // A ClawHub package is verified only once installed: its files must match the
  // release commit's hashes, or the update is rolled back like a rejected one.
  const mismatch = st.kind === "clawhub" ? await installedMismatch(st.files) : undefined;
  if (mismatch) return rollback(st, `the installed ClawHub package differs from the release (${mismatch})`, log, true);
  log.info(`[egbe] update to ${st.toVersion} is live`);
  return writeState({ phase: "idle", lastResult: `updated ${st.fromVersion} -> ${st.toVersion}`, lastSuccessAt: new Date().toISOString() });
}

Technical Analysis

The updater verifies release information and files from a Git commit, but the ClawHub update branch subsequently installs a separately distributed package by version. The verified Git content is not the artifact passed to installation.

The plugin compares the installed package against the release hashes only after LOADED_VERSION matches the target version. At that point, the replacement module has already loaded and can execute initialization and registration code with gateway permissions.

Rollback is therefore a ...[truncated 2368 chars]

SkillSpector findings are pending for this release.

Endor dependency reachability

Checked Nov 18, 2026 · 3:07 PM UTC

No reachable vulnerabilities reported

Endor dependency scan completed.