Back to skill

Security audit

Wadiz DM Pipeline

Security checks for vulnerabilities and agentic risk

Overview

This is a local Wadiz crowdfunding crawler and DM CSV generator, with some disclosure gaps around Anthropic API use and unsupported Tumblbug wording.

Install only if you are comfortable running a local browser crawler, creating CSV outreach files, and, when ANTHROPIC_API_KEY is configured, sending scraped project text to Anthropic for DM generation. Expect Wadiz support only despite the Tumblbug wording, review generated DMs before sending, and use an isolated environment with pinned dependencies where possible.

Vulnerability Patterns
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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
Findings (2)

T08 · Insecure Dependencies

Warning
Location
SKILL.md:40
Finding

Unpinned third-party dependencies and browser artifacts

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:40 Vulnerability Type: Unpinned dependency installation Risk Level: Medium

Vulnerable Code Snippet

bash
- Playwright required: `pip install playwright && python -m playwright install chromium`

Related installation guidance also appears in scripts/pipeline.py:21-35:

python
try:
    from playwright.sync_api import sync_playwright
except ImportError:
    print("❌ Playwright is not installed. Run the following commands and try again:")
    print("   pip install playwright")
    print("   python -m playwright install chromium")
    sys.exit(1)

try:
    import anthropic
except ImportError:
    print("❌ anthropic is not installed. Run the following command and try again:")
    print("   pip install anthropic")
    sys.exit(1)

Technical Analysis

The project instructs users to install playwright and anthropic without fixed versions, package hashes, or a lock file. Playwright also downloads a Chromium artifact whose version and integrity are not independently pinned or verified by this project.

Consequently, installation depends on mutable package-index and browser-download state at the time the commands are executed. Although there is no evidence that the currently named packages are malicious, a compromised upstream account, release, transitive dependency, package index, or browser distribution channel could introduce executable code after the skill has been audited.

Attack Path

  1. An attacker compromises an upstream package, transitive dependency, maintainer account, package distribution channel, or browser artifact source.
  2. The attacker publishes or substitutes a malicious release.
  3. A user follows the documented unpinned installation commands.
  4. Package resolution or Playwright's installer retrieves the compromised component.
  5. Malicious installation or runtime code executes under the account ...[truncated 662 chars]
Remediation
View remediation

Remediation Suggestions

  1. Add a reviewed requirements or lock file with exact versions for playwright, anthropic, and relevant transitive dependencies.
  2. Generate and verify package hashes, using an installation process such as pip install --require-hashes -r requirements.txt.
  3. Configure an explicit trusted package index or an internally controlled dependency mirror.
  4. Pin the expected Playwright and Chromium versions and document how the browser artifact's provenance and integrity are verified.
  5. Run dependency installation and the pipeline as an unprivileged user in an isolated virtual environment.
  6. Add automated dependency scanning and a controlled process for reviewing version updates before changing the lock file.

T09 · Insecure Skill Coding Practices

Warning
Location
scripts/pipeline.py:132
Finding

Indirect prompt injection through untrusted scraped project content

Content
View full analysis

Vulnerability Details

File Location: scripts/pipeline.py:132-178 Vulnerability Type: Indirect prompt injection Risk Level: Medium

Vulnerable Code Snippet

python
def get_project_detail(page, url: str) -> str:
    """Extract key content from the project detail page."""
    try:
        page.goto(url, wait_until="networkidle", timeout=20000)
        time.sleep(1)
        texts = []
        for sel in ["h1", ".reward-intro", "[class*='intro']", "[class*='summary']", "[class*='description']"]:
            els = page.query_selector_all(sel)
            for el in els[:2]:
                t = el.inner_text().strip()
                if len(t) > 10:
                    texts.append(t[:300])
        return " | ".join(texts[:3])
    except Exception:
        return ""


def generate_dm(title: str, description: str) -> str:
    """Generate a personalized DM through the Anthropic API."""
    if not ANTHROPIC_API_KEY:
        hook = f"{title} has an interesting product direction."
        return DM_TEMPLATE.format(project_name=title, hook=hook)

    client = anthropic.Anthropic(api_key=ANTHROPIC_API_KEY)
    prompt = f"""Review the following crowdfunding project and write a short DM message for its creator.

Project name: {title}
Project content: {description[:500]}

DM requirements:
- No more than 100 words
- The first line must identify one project characteristic and say that it prompted the contact
- Never mention AI
- Do not boast
- Include the specified service offer
- End with the specified soft call to action
- Use natural and respectful language

Output only the DM text without additional explanation."""

    try:
        msg = client.messages.create(
            model="claude-haiku-4-5",
            max_tokens=300,
            messages=[{"role": "user", "content": prompt}],
        )
        return msg.content[0].text.strip()
    except Exception as 
...[truncated 2582 chars]
Remediation
View remediation

Remediation Suggestions

  1. Put immutable behavioral rules in a system message rather than combining all instructions and untrusted data in one user message.
  2. Delimit project fields using an explicit structured format such as JSON or XML and state that all content inside those fields is untrusted data that must never be followed as instructions.
  3. Normalize input and detect instruction-like phrases, suspicious URLs, or attempts to override the generation policy before invoking the model.
  4. Validate generated messages against an allowlisted output schema and enforce required content, length limits, and the expected closing text.
  5. Reject or flag generated output containing unexpected URLs, contact details, system-like instructions, unrelated promotions, or calls to action outside the approved template.
  6. Preserve human review before any DM is sent; do not add automatic delivery without stronger isolation and validation.
  7. Consider generating only the personalized hook with the model and assembling the rest of the message from trusted static template text.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • 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 (6)

Tp4

High
Category
MCP Tool Poisoning
Confidence
96% confidence
Finding

The skill description understates or misstates actual behavior: it claims Wadiz/Tumblbug support, while the finding indicates Tumblbug is not implemented and external web access plus Anthropic API calls occur without declared permissions. Behavior/description mismatches are dangerous because they defeat informed consent and can cause unreviewed outbound data flows to third parties, especially if scraped project data or prompts are sent to an external LLM service.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
92% confidence
Finding

The skill advertises executable behavior that implies access to local environment state and writing output files, but it does not declare any tool scope or permissions boundaries. This is dangerous because users and the runtime cannot clearly understand or constrain what resources the skill may access, increasing the chance of unintended file writes, secret exposure, or over-privileged execution.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

Broad trigger phrases like general requests to scrape projects or generate DMs create weak activation boundaries, making accidental or overly frequent invocation more likely. In a skill that performs crawling and disk writes, ambiguous triggers can lead to unintended external requests, local artifact creation, and surprise execution without the user realizing the full operational scope.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The prompt explicitly requires 자연스러운 한국어, which imposes a specific language on generated content. Under the policy, language constraints should either be user-selectable or clearly documented as justified for a region-specific skill; this file does not present a user opt-in or explicit justification mechanism.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The script sends scraped project descriptions to Anthropic for DM generation without any explicit in-code consent, disclosure, or data-handling notice. Even if the source data is public, this creates an undisclosed third-party transfer and can violate user expectations, platform policies, or internal privacy/compliance requirements when run as an automation pipeline.

Content

No source excerpt is available for this finding.

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
87% confidence
Finding

The skill text does not prominently warn that it crawls external websites and writes results to local disk, which reduces transparency around side effects. While this is less severe than hidden secret exfiltration, it is still risky because users may trigger network activity or create local CSVs containing scraped or generated outreach content without realizing it.

Content

No source excerpt is available for this finding.