Back to skill

Security audit

Wolverine — Self-Healing Process Manager

Security checks for vulnerabilities and agentic risk

Overview

The skill has a clear self-healing purpose, but it relies on an unpinned remote install and grants a tool authority to modify code, send crash diagnostics to AI services, and keep persistent backups or memory.

Install only in a staging or low-risk workspace, preferably with a pinned package version and a low-privilege account. Review what crash data may leave the machine, where backups and memory are stored, and whether fixes require your approval before they are applied.

Vulnerability Patterns
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • 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
  • 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)

T03 · Remote Payload Retrieval and Execution

Error
Location
SKILL.md:43
Finding

Mutable Remote Package Is Downloaded and Executed Through npx

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:43
Vulnerability Type: Remote payload retrieval and supply-chain risk
Risk Level: High

Evidence

bash
npx wolverine-ai@latest --setup-claw

The package is also declared without an exact version:

yaml
install:
  - kind: node
    package: wolverine-ai
    bins:
      - wolverine
      - wolverine-claw

Technical Analysis

The documented setup command instructs users to download and execute the package version associated with the mutable latest npm tag. The effective executable payload can therefore change after this Skill has been reviewed.

The project contains only SKILL.md; it does not include the package implementation, a lockfile, an integrity hash, or vendored source through which the downloaded code could be verified. Any lifecycle scripts or transitive dependencies associated with the current package release may execute under the invoking user's account.

This behavior exceeds the minimum privileges needed for deterministic installation because an audited, exact package version could be used instead of a mutable release tag.

Attack Path

  1. An attacker compromises the npm publisher account, the package, or one of its dependencies.
  2. The attacker publishes a malicious release and assigns it to the latest tag.
  3. A user follows the documented Quick Start command.
  4. npx downloads and executes the attacker-controlled release.
  5. The payload runs with the user's permissions and may access the workspace, environment variables, API credentials, files, and processes available to that account.

Impact Assessment

Successful exploitation could result in arbitrary code execution with the privileges of the user running npx. The accessible scope may include the OpenClaw workspace, ANTHROPIC_API_KEY, optional OPENAI_API_KEY, user-readable files, network access, and any processes the account is permitted t ...[truncated 221 chars]

Remediation
View remediation

Remediation Suggestions

  • Replace @latest with an exact, reviewed version.
  • Commit a lockfile containing resolved transitive dependency versions and integrity metadata.
  • Verify npm package integrity and publisher provenance before execution.
  • Disable or explicitly review package lifecycle scripts where practical.
  • Prefer installing the package separately and invoking a verified local binary rather than combining retrieval and execution in one command.
  • Publish the reviewed implementation with the Skill or link it to an immutable source revision.
  • Run installation and setup under a dedicated, low-privilege account without access to unrelated credentials or production data.
  • Establish an update process that audits each new package version before changing the pinned version.

other

Warning
Location
SKILL.md:31
Finding

Crash and Diagnostic Information Is Transmitted to External AI Providers

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:31-35
Vulnerability Type: External AI data disclosure
Risk Level: Medium

Evidence

markdown
Wolverine watches your OpenClaw process. When it crashes, Wolverine:

1. Captures the error (crash or caught 500)
2. Diagnoses it with AI (Anthropic or OpenAI)
3. Proposes a code fix

The Skill also confirms that data is sent in AI calls:

markdown
| Secret Redactor | Scrubs API keys from all AI calls, logs, and memory |

Technical Analysis

AI-based crash diagnosis requires diagnostic information to cross the local trust boundary and reach Anthropic or OpenAI. Error messages and associated diagnostic context may contain stack traces, source excerpts, local paths, request data, configuration values, tokens, or other confidential operational information.

This network behavior is relevant to the declared AI diagnosis function and is not evidence of covert exfiltration. However, the artifact does not disclose the exact fields transmitted, retention controls, endpoints, redaction algorithm, or whether transmission fails closed when redaction cannot be completed.

The Skill claims that API keys are scrubbed, but the redaction implementation is contained in an external package and was not available for audit. Its effectiveness against unknown secret formats, credentials embedded in URLs, multiline values, source-code secrets, and non-key personal or business data therefore cannot be verified.

Attack Path

  1. OpenClaw crashes or returns a caught server error.
  2. The error or related diagnostic context contains confidential data.
  3. Wolverine captures that context for AI diagnosis.
  4. A value not recognized by the secret-redaction rules remains in the diagnostic payload.
  5. The payload is transmitted to Anthropic or OpenAI.
  6. The information is exposed outside the local workspace and becomes subject to the provider's processing and retenti ...[truncated 511 chars]
Remediation
View remediation

Remediation Suggestions

  • Require explicit, informed opt-in before transmitting diagnostic data.
  • Document the providers, endpoints, transmitted fields, retention policy, and applicable privacy boundaries.
  • Minimize payloads to the smallest context required for diagnosis.
  • Use allowlist-based serialization so only approved diagnostic fields can leave the host.
  • Redact API keys, authorization headers, cookies, credentials in URLs, environment variables, personal data, and organization-specific secret formats.
  • Fail closed and request human review if redaction cannot be completed reliably.
  • Display or log a local preview of the exact redacted payload before transmission.
  • Provide a local-only mode that performs restart and operational recovery without external AI calls.
  • Avoid retaining unredacted diagnostic data in logs, backups, or semantic memory.
  • Test the redactor against encoded, multiline, nested, and previously unseen secret formats.
Vulnerability Patterns
  • Rogue AgentSelf-Modification, Session Persistence
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (2)

Session Persistence

Medium
Category
Rogue Agent
Confidence
87% confidence
Finding

The skill explicitly stores backups and semantic memory across runs, meaning crash data, code snapshots, and prior fixes may persist outside the project directory. In a self-healing tool that captures errors and sends diagnostics to AI services, this persistence can retain sensitive source code, stack traces, secrets, or operational metadata longer than expected, increasing exposure if the host is shared or the backup area is not adequately protected.

Content

Scanner excerpt · SKILL.md (reported line 26)May include surrounding context.

md
# Wolverine — Self-Healing Process Manager

> **EXPERIMENTAL**: This skill wraps your OpenClaw gateway in a supervised repair and security layer. Do not use in workspaces with critical production data until you have tested in a staging environment. All file modifications create backups first and can be rolled back.

## What It Does

Rp1

Medium
Category
MCP Rug Pull
Confidence
95% confidence
Finding

The Quick Start command uses npx wolverine-ai@latest --setup-claw, which fetches and executes the newest package version at runtime rather than a reviewed, pinned release. This creates a supply-chain risk: if the npm package is compromised or a bad release is published, users may immediately execute untrusted code with their local permissions.

Content

No source excerpt is available for this finding.