Back to skill

Security audit

claim-audit

Security checks for vulnerabilities and agentic risk

Overview

ClawScan marked this skill suspicious because it is described as a read-only audit tool, but scanners report that it can run package-provided code and write a results file.

Consider not installing until you have reviewed these execution and file-writing behaviors and can run the skill in an isolated environment. The scan is marked partial: although it reports all 12 listed components scanned, some referenced paths were missing, so the evidence is incomplete. A separate AIG scan completed with no findings, but that does not resolve the concerns reported by Skillspector.

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
Vulnerability Patterns
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (10)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The unconditional read-only assurance is contradicted by the supplied self-test runner, which opens tests/result.json for writing. Auditing this package's test claims can therefore modify the evidence being audited or overwrite an existing result artifact without the user's informed consent.

Content

No source excerpt is available for this finding.

Skill declares no tool scope ('permissions' or 'allowed-tools') but code capabilities were detected: file_read, file_write, shell.

Medium
Category
MCP Least Privilege
Confidence
70% confidence
Finding

Without declared permissions the skill's intent is opaque and cannot be validated.

Content

No source excerpt is available for this finding.

The workflow runs the audited package's code without warning that execution can have side effects, while describing the audit as read-only.

Medium
Category
Output Handling
Confidence
99% confidence
Finding

The workflow directs execution of code supplied by the package being audited while presenting the operation as read-only. An adversarial package can use its test runner to read secrets, modify files, or access the network with the auditor's privileges; a timeout does not prevent these side effects.

Content

No source excerpt is available for this finding.

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · demo/run_demo.py (reported line 29)May include surrounding context.

python
root = tempfile.mkdtemp(prefix="demo-")
    w(root, "SKILL.md", "本包自检 9/9 PASS,共 12 个文件。\n")
    w(root, "tests/run_golden.py", 'print("PASS T1")\nprint("--- 1/2 PASS ---")\n')
    r = subprocess.run([sys.executable, AUDIT, "check", root],
                       capture_output=True, text=True, encoding="utf-8")
    print(r.stdout)
    print("退出码:", r.returncode, "(1 = 有声明对不上)")

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
99% confidence
Finding

The selected runner is Python code controlled by the audited package and executes directly with the auditor's privileges and inherited environment. Although the argument list avoids shell injection, it does not prevent the runner itself from stealing credentials, modifying host files, or spawning additional processes.

Content

Scanner excerpt · scripts/audit.py (reported line 103)May include surrounding context.

python
if runner is None:
        return None
    try:
        r = subprocess.run([sys.executable, runner], cwd=root, capture_output=True,
                           text=True, encoding="utf-8", timeout=180)
    except Exception:  # noqa: BLE001
        return None

Unparseable test results are reported as mismatches rather than unverifiable claims.

Medium
Category
Excessive Agency
Confidence
98% confidence
Finding

When no supported result is found, the function fabricates a 0/0 result instead of preserving uncertainty, producing unsupported mismatch verdicts. A 0/0 PASS claim could even be accepted without parsed evidence, undermining the integrity of this auditing tool's conclusions.

Content

No source excerpt is available for this finding.

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
98% confidence
Finding

The reviewer is selected solely by its presence at a predictable sibling or ancestor-relative path, without establishing its provenance. An attacker who supplies the surrounding directory layout can plant that script for execution, and the --run option may additionally execute audited-package code.

Content

Scanner excerpt · scripts/audit.py (reported line 128)May include surrounding context.

python
if not rv:
        return None
    try:
        r = subprocess.run([sys.executable, rv, root, "--run"], capture_output=True,
                           text=True, encoding="utf-8", timeout=180)
    except Exception:  # noqa: BLE001
        return None

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · tests/run_golden.py (reported line 49)May include surrounding context.

python
def run(args, cwd=None):
    p = subprocess.run([sys.executable, AUDIT] + args, capture_output=True,
                       text=True, encoding="utf-8", timeout=120, cwd=cwd)
    return p.returncode, (p.stdout or "") + (p.stderr or "")

The self-test runner persistently writes tests/result.json despite the manifest's read-only claim.

Low
Category
Excessive Agency
Confidence
99% confidence
Finding

The runner unconditionally opens tests/result.json in write mode, creating or truncating a persistent file in the package. When invoked during an audit, this changes the audited evidence despite the read-only assurance and can destroy an existing result artifact.

Content

No source excerpt is available for this finding.

The test runner silently creates or overwrites tests/result.json.

Low
Category
Output Handling
Confidence
98% confidence
Finding

The result artifact is written without an output option or overwrite confirmation, making persistent mutation an implicit side effect of running tests. This is particularly problematic in an evidence-verification workflow advertised as read-only.

Content

No source excerpt is available for this finding.