Back to skill

Security audit

龙虾数字人

Security checks for vulnerabilities and agentic risk

Overview

This skill does perform digital-human video generation, but it handles credentials and sensitive media in ways that create significant review-worthy risk.

Review this carefully before installing. Use only on networks and accounts where you are comfortable sending API tokens, scripts, voice recordings, videos, images, and task results to the easyclaw.bar platform. The current artifacts default to HTTP, so credentials and media may be exposed in transit. Avoid enabling the async watcher until credentials are no longer written into cron prompts or command-line arguments, and rotate any tokens previously used with this version.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • System PersistenceInstalls backdoors, hooks, services, or scheduled tasks that survive the run
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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
Findings (3)

T09 · Insecure Skill Coding Practices

Error
Location
scripts/client.py:14
Finding

Authentication Credentials and Biometric Media Transmitted over Plaintext HTTP

Content
View full analysis
Remediation
View remediation

T06 · System Persistence

Error
Location
scripts/schedule_task_watch.py:216
Finding

API Secrets Embedded in Persistent Cron Prompts, Process Arguments, and Script Output

Content
View full analysis
list[str]: args: list[str] = [] base_url = os.environ.get("CHANJING_PLATFORM_BASE_URL", "").strip() platform_token = os.environ.get("CHANJING_PLATFORM_API_TOKEN", "").strip() api_key = os.environ.get("CHANJING_PLATFORM_API_KEY", "").strip() api_secret = os.environ.get("CHANJING_PLATFORM_API_SECRET", "").strip() if base_url: args.extend(["--base-url", base_url]) if platform_token: args.extend(["--api-token", platform_token]) return args if api_key and api_secret: args.extend(["--api-key", api_key, "--api-secret", api_secret]) return args ``` ```python def _shell_command(script_path: Path, *, task_id: int, task_kind: str, label: str, watch_key: str, job_id: str | None) -> str: args = [ sys.executable, str(script_path), "--task-id", str(task_id), "--task-kind", task_kind, "--watch-key", watch_key, ] if label: args.extend(["--label", label]) if job_id: args.extend(["--job-id", job_id]) args.extend(_watcher_auth_args()) return " ".join(shlex.quote(part) for part in args) ``` ```python def _build_prompt(script_path: Path, *, task_id: int, task_kind: str, label: str, watch_key: str, job_id: str | None) -> str: command = _shell_command( script_path, task_id=task_id, task_kind=task_kind, label=label, watch_key=watch_key, job_id=job_id, ) return "\n".join( [ "Check one async Chanjing platform task.", ...[truncated 3189 chars]
Remediation
View remediation

T01 · Skill Instruction Hijacking

Error
Location
scripts/schedule_task_watch.py:251
Finding

Untrusted Remote Task Data Is Relayed Verbatim by Persistent Agent Turns

Content
View full analysis
str: command = _shell_command( script_path, task_id=task_id, task_kind=task_kind, label=label, watch_key=watch_key, job_id=job_id, ) return "\n".join( [ "Check one async Chanjing platform task.", "Run this exact command in the terminal:", command, "If the command prints exactly NO_REPLY, reply with exactly NO_REPLY and nothing else.", "Otherwise reply with the command output verbatim and nothing else.", ] ) ``` ```python def build_success_message(task_payload: dict, task_kind: str, label: str) -> str: task = task_payload.get("task") or {} result_payload = task_payload.get("result_payload") or {} resource_label = label or result_payload.get("name") or result_payload.get("title") or task_kind lines = [ f"{task_kind} task completed: {resource_label}", f"Task ID: #{task.get('id')}", f"Status: {task.get('status')}", ] upstream_resource_id = task.get("upstream_resource_id") if upstream_resource_id: lines.append(f"Resource ID: {upstream_resource_id}") media = collect_media(result_payload)[:3] for entry in media: media_type = entry.get("type") url = entry.get("url") if not url: continue if media_type == "image": lines.append(f"![{entry.get('label') or resource_label}]({url})") elif media_type == "video": lines.append(f"Video: {url}") elif media_type == "audio": lines.append(f"Aud ...[truncated 3719 chars]
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • 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
Findings (35)

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The skill includes cron/job creation, watcher management, local persistence under user directories, environment credential use, and long-running task monitoring, none of which are disclosed in the high-level description. Hidden background scheduling and local state creation can surprise users, expand persistence on the host, and increase the blast radius if the skill or its scripts are misused.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The skill includes cron/job creation, watcher management, local persistence under user directories, environment credential use, and long-running task monitoring, none of which are disclosed in the high-level description. Hidden background scheduling and local state creation can surprise users, expand persistence on the host, and increase the blast radius if the skill or its scripts are misused.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The skill includes cron/job creation, watcher management, local persistence under user directories, environment credential use, and long-running task monitoring, none of which are disclosed in the high-level description. Hidden background scheduling and local state creation can surprise users, expand persistence on the host, and increase the blast radius if the skill or its scripts are misused.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The skill includes cron/job creation, watcher management, local persistence under user directories, environment credential use, and long-running task monitoring, none of which are disclosed in the high-level description. Hidden background scheduling and local state creation can surprise users, expand persistence on the host, and increase the blast radius if the skill or its scripts are misused.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
95% confidence
Finding

The skill includes cron/job creation, watcher management, local persistence under user directories, environment credential use, and long-running task monitoring, none of which are disclosed in the high-level description. Hidden background scheduling and local state creation can surprise users, expand persistence on the host, and increase the blast radius if the skill or its scripts are misused.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

The skill includes cron/job creation, watcher management, local persistence under user directories, environment credential use, and long-running task monitoring, none of which are disclosed in the high-level description. Hidden background scheduling and local state creation can surprise users, expand persistence on the host, and increase the blast radius if the skill or its scripts are misused.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The skill includes cron/job creation, watcher management, local persistence under user directories, environment credential use, and long-running task monitoring, none of which are disclosed in the high-level description. Hidden background scheduling and local state creation can surprise users, expand persistence on the host, and increase the blast radius if the skill or its scripts are misused.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
89% confidence
Finding

The skill includes cron/job creation, watcher management, local persistence under user directories, environment credential use, and long-running task monitoring, none of which are disclosed in the high-level description. Hidden background scheduling and local state creation can surprise users, expand persistence on the host, and increase the blast radius if the skill or its scripts are misused.

Content

No source excerpt is available for this finding.

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The skill includes cron/job creation, watcher management, local persistence under user directories, environment credential use, and long-running task monitoring, none of which are disclosed in the high-level description. Hidden background scheduling and local state creation can surprise users, expand persistence on the host, and increase the blast radius if the skill or its scripts are misused.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The code probes local installation directories and executes external OpenClaw/EasyClaw/ClawX commands through subprocess, which is outside the obvious scope of video generation. In context, this broadens the attack surface to local process execution and environment-dependent behavior with little justification from the skill metadata.

Content

No source excerpt is available for this finding.

Env Variable Harvesting

High
Category
Data Exfiltration
Confidence
60% confidence
Finding

Code enumerates, copies, or searches environment variables for secrets. Bulk environment access can collect credentials unrelated to the skill's stated purpose.

Content

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

python
last_result: subprocess.CompletedProcess[str] | None = None
    for base_command, extra_env in _openclaw_command_candidates():
        command = [*base_command, *args]
        env = os.environ.copy()
        env.update(extra_env)
        try:
            result = subprocess.run(command, capture_output=True, text=True, check=False, env=env)

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

This file implements creation, editing, listing, and removal of OpenClaw cron jobs, which is unrelated to the declared digital-human video generation function. Such hidden scheduler control materially expands the skill's capabilities and can establish persistence or autonomous follow-on actions without matching user expectations.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
95% confidence
Finding

The skill advertises executable capabilities including environment access, file read/write, network access, and shell execution, but does not declare any explicit tool scope restrictions such as permissions or allowed-tools. This weakens least-privilege controls and makes it easier for the skill to access sensitive local data, invoke arbitrary commands, or reach external services beyond what users would reasonably infer from the description.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The skill processes potentially sensitive user media, scripts, voices, and training assets via an external platform API but does not warn users that their content will be transmitted off-platform. For a media-generation skill involving personal likeness and voice training, this omission is significant because users may unknowingly upload biometric/personal data to a third party.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The skill hard-codes a specific external URL and instructs the agent to direct users through a fixed Chinese-language onboarding flow without asking for consent, region preference, or locale suitability. This can steer users to an unexpected third-party service, reduce informed choice, and facilitate phishing-style trust in a single externally controlled domain.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The reference documents authenticated calls and data-submitting endpoints for training and video synthesis, but it does not prominently warn that user-provided media, scripts, and related metadata will be transmitted to an external platform service. In a skill that handles voice, video, and digital human training, this omission can cause users or downstream integrators to send sensitive biometric, personal, or proprietary content without clear notice or consent expectations.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The setup instructions direct users to log in, enter upstream API credentials, and generate/copy an API token over plaintext HTTP endpoints. Because HTTP provides no transport encryption or server authentication, credentials and tokens can be intercepted or modified by network attackers, enabling account takeover and unauthorized API use. In this skill context, the risk is elevated because the document explicitly normalizes handling multiple sensitive secrets through insecure URLs.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The client reads arbitrary local files and uploads their full contents to a remote service over plain HTTP, which creates a real confidentiality risk. In this skill context, uploading user media is expected functionality, but the code provides no in-code consent boundary or destination safety guarantees, so sensitive local files could be exfiltrated if the caller passes an unintended path or if traffic is intercepted.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
86% confidence
Finding

This code accepts base URL and API credentials via command-line arguments and immediately configures the client with them, but there is no visible warning, comment, or docstring explaining that sensitive authentication material is being handled. Under the code-file criteria, access to sensitive credentials should have some form of disclosure unless clearly warned elsewhere in the skill description, which is not visible in this file.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The parser description is hardcoded in Chinese ("获取当前用户的音色列表"), which indicates a fixed language choice in the skill's user-facing text. The file does not offer language selection or document that the skill is intentionally region- or locale-specific.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
86% confidence
Finding

The script executes external commands and is used to schedule autonomous watcher behavior without any explicit confirmation or warning to the user. In a skill that is supposed to generate digital-human videos, silently creating recurring jobs is more dangerous because it introduces persistence and background execution outside normal user expectations.

Content

No source excerpt is available for this finding.

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
89% confidence
Finding

The code launches external binaries discovered from the local environment and installation paths, giving this skill the ability to execute arbitrary local programs named or pointed to as OpenClaw/EasyClaw/ClawX. Although subprocess is used without shell=True, the target executable can still be attacker-controlled via PATH or OPENCLAW_BIN, making this a real command-execution surface in a skill whose stated purpose does not require local scheduler control.

Content

Scanner excerpt · scripts/schedule_task_watch.py (reported line 131)May include surrounding context.

python
env = os.environ.copy()
        env.update(extra_env)
        try:
            result = subprocess.run(command, capture_output=True, text=True, check=False, env=env)
        except OSError:
            continue
        if result.returncode == 0:

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The script reads platform credentials from environment variables and embeds them into arguments used to construct scheduled watcher commands and prompts. This exceeds the declared skill purpose and creates a pathway for secrets to be exposed to downstream tooling, logs, prompts, or other scheduled execution contexts.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

Credentials are sourced from environment variables and then appended as CLI arguments, which are often exposed through process listings, job metadata, logs, or diagnostic output. The lack of disclosure or warning worsens the risk because users are not informed that secrets may be propagated into observable execution paths.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The generated prompt instructs the scheduled agent to return command output verbatim, and the command may include authentication arguments from environment-derived credentials. This creates a direct exfiltration path where secrets, sensitive API responses, or internal diagnostics can be echoed back unchanged through the scheduler channel.

Content

No source excerpt is available for this finding.