Back to skill

Security audit

only for test

Security checks for vulnerabilities and agentic risk

Overview

This document extraction skill is mostly purpose-aligned, but its read-only framing conflicts with broad command/write permissions, unsafe temporary PDF image files, and under-described OCR data handling.

Install only if you are comfortable granting broad local command and file-write authority to a document extraction skill. Avoid using it on confidential scans or screenshots unless you know how the OCR MCP service handles data, and prefer running dependency installs in a dedicated virtual environment. Scanned PDFs may leave page images in /tmp unless the workflow is fixed to use private temporary files and cleanup.

Vulnerability Patterns
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • 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
Findings (4)

T08 · Insecure Dependencies

Warning
Location
SKILL.md:59
Finding

Unpinned Runtime Dependency Installation

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:59-71
Vulnerability Type: Supply-chain exposure through unpinned Python dependencies
Risk Level: Medium

bash
pip install python-docx openpyxl pymupdf

Technical Analysis

The Skill instructs users to install three packages without exact version pins, integrity hashes, a lockfile, or an isolated environment. Although minimum versions are documented separately, the installation command resolves the latest matching packages from the configured Python package index.

This makes the effective code installed at runtime mutable after the Skill has been reviewed. Security depends on the integrity of the package registry, index configuration, package maintainers, and all transitive dependencies.

Attack Path

  1. An attacker compromises a listed package, one of its transitive dependencies, or the package index used by the environment.
  2. The victim follows the documented installation command.
  3. pip resolves and installs the attacker-controlled release.
  4. Malicious installation hooks or imported package code execute under the victim's account when the dependency is installed or the Skill processes a document.

Impact Assessment

Successful exploitation could execute arbitrary code with the privileges of the user running pip or invoking the Skill. This may expose local documents, credentials accessible to that account, and other files within the account's permission scope. If installation is performed with elevated privileges, the impact could extend to system-wide compromise.

Remediation
View remediation

Remediation Suggestions

  • Pin every direct and transitive dependency to a reviewed version.
  • Use a lockfile or constraints file generated from a trusted environment.
  • Require package hashes, such as through pip install --require-hashes.
  • Install dependencies inside a dedicated, non-privileged virtual environment.
  • Configure an approved package index and disable untrusted supplemental indexes.
  • Add automated dependency vulnerability and integrity scanning.
  • Upgrade packages through a controlled review process rather than resolving mutable versions during Skill execution.

T05 · Unauthorized Access and Privilege Escalation

Error
Location
SKILL.md:17
Finding

Read-Only Skill Grants General Command and File-Write Capabilities

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:17
Vulnerability Type: Excessive tool permissions and violation of least privilege
Risk Level: High

yaml
allowed-tools: Read, Grep, Glob, Write, Bash, mcp__zai-mcp-server__extract_text_from_screenshot

Technical Analysis

The Skill is defined as a read-only document extraction utility, but its configuration authorizes unrestricted Write and Bash tools. General shell execution and file modification are substantially broader than the legitimate requirements of reading documents and returning extracted Markdown.

Because documents are attacker-controlled input, their extracted contents may contain prompt-injection instructions. If such content is mistakenly treated as agent instructions, the pre-authorized Bash and Write capabilities provide a direct path from untrusted document text to local command execution or file modification.

Attack Path

  1. An attacker creates a supported document containing instructions intended to manipulate the agent after extraction.
  2. A victim asks the Skill to process the document.
  3. The parser or OCR service returns the attacker-controlled instructions as document text.
  4. The agent incorrectly treats those instructions as operational commands rather than untrusted data.
  5. The agent invokes Bash or Write to execute commands, alter files, or access resources unrelated to document extraction.

Impact Assessment

Exploitation could execute commands and modify files with the privileges of the agent process. The accessible scope may include the project workspace, user-owned files, environment variables, credentials available to subprocesses, and network resources reachable from the host. No privilege elevation beyond the agent process's operating-system account is demonstrated, but the granted capabilities exceed the Skill's stated purpose.

Remediation
View remediation

Remediation Suggestions

  • Remove general Bash and unrestricted Write permissions.
  • Replace them with narrowly scoped document-parsing operations.
  • If temporary writes are necessary, permit writes only within a private temporary directory.
  • Prohibit command execution derived from document contents or OCR results.
  • Explicitly mark all extracted text as untrusted data that cannot modify Skill instructions.
  • Require confirmation for operations outside extraction and formatting.
  • Run parsers in a sandbox with restricted filesystem access, no unnecessary network access, resource limits, and a non-privileged account.

T09 · Insecure Skill Coding Practices

Warning
Location
SKILL.md:117
Finding

Predictable Temporary Filenames for Rendered PDF Pages

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:117-125
Vulnerability Type: Unsafe temporary-file creation and sensitive-data retention
Risk Level: Medium

python
for i, page in enumerate(doc, 1):
    text = page.get_text().strip()
    if text:
        # Text-based PDF extraction
    else:
        # Render scanned PDF page for OCR
        pix = page.get_pixmap(dpi=150)
        pix.save(f"/tmp/pdf_page_{i}.png")
        # Invoke the configured MCP OCR tool

Technical Analysis

Rendered PDF pages are written to predictable global paths such as /tmp/pdf_page_1.png. The filenames contain no random component, and the instructions do not create a private directory, verify file ownership, reject symbolic links, apply restrictive permissions, or remove the images after OCR.

In a shared environment, concurrent processes can collide with these filenames. A local attacker may pre-create a file or symbolic link at the expected path. Depending on the image library's file-opening behavior and operating-system permissions, this could cause denial of service, overwrite an attacker-selected file, or expose rendered document contents. Even without active exploitation, omitted cleanup may leave sensitive page images available after processing.

Attack Path

  1. A local attacker predicts that the first scanned page will be written to /tmp/pdf_page_1.png.
  2. The attacker monitors that path or pre-creates a conflicting file or symbolic link.
  3. The victim processes a scanned PDF.
  4. The Skill writes the rendered page to the predictable path.
  5. The attacker reads the retained page, disrupts extraction, or attempts to redirect the write to another location permitted by the victim's account.

Impact Assessment

The primary impact is disclosure of potentially confidential PDF content to other local users or processes. Additional impacts may include extraction failures, cross-process data corruption, ...[truncated 152 chars]

Remediation
View remediation

Remediation Suggestions

  • Create a private temporary directory with tempfile.TemporaryDirectory.
  • Generate unpredictable filenames instead of using page numbers in a shared directory.
  • Set restrictive permissions so only the current process or user can read rendered pages.
  • Prevent symbolic-link following and use exclusive file creation where supported.
  • Delete every rendered page immediately after OCR and guarantee cleanup with a finally block or context manager.
  • Avoid retaining rendered images in logs, caches, or output artifacts.
  • Isolate each extraction request in a separate temporary directory to prevent concurrent-job collisions.

other

Warning
Location
SKILL.md:44
Finding

Document Images Are Sent to an OCR Integration Without a Defined Trust Boundary

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:44-55
Vulnerability Type: Potential sensitive-document disclosure to an externally managed OCR integration
Risk Level: Medium

text
Tool: mcp__zai-mcp-server__extract_text_from_screenshot

Parameters:
  image_source: image path or URL
  prompt: extraction instruction
  programming_language: optional code language

Technical Analysis

The Skill requires an MCP OCR tool for images and scanned PDF pages. The documentation does not define whether processing is local or remote, identify the resolved service endpoint, state retention and logging policies, describe authentication controls, or require informed user consent before document images cross that integration boundary.

MCP is an integration mechanism rather than a guarantee of local processing. Supplying a local path may permit the integration to read the referenced image, while remote implementations may transmit the image or extracted content to another service. Confidentiality therefore depends on infrastructure that is not described or constrained by the Skill.

Attack Path

  1. A victim requests extraction from a confidential image or scanned PDF.
  2. For a scanned PDF, the Skill renders the page to an image.
  3. The image path or image content is supplied to the configured MCP OCR integration.
  4. The integration processes, transmits, logs, or retains the content according to policies not disclosed in the Skill.
  5. An operator, compromised integration, or downstream service may gain access to the document content.

Impact Assessment

The potential impact is disclosure of all visible content in submitted images or scanned PDF pages, including personal data, business records, source code, credentials shown in screenshots, or regulated information. The exact destination and retention scope cannot be determined from the audited file, so external transmission is a trust-boundary risk rather ...[truncated 49 chars]

Remediation
View remediation

Remediation Suggestions

  • State whether OCR is performed locally or remotely and identify the responsible service.
  • Display the resolved endpoint and data-handling policy before processing.
  • Obtain explicit user consent before transmitting document images outside the local environment.
  • Document retention, logging, encryption, authentication, and deletion policies.
  • Provide a local-only OCR option for confidential material.
  • Minimize transmitted data by cropping or redacting unnecessary regions.
  • Reject public URLs or untrusted OCR destinations unless explicitly authorized.
  • Ensure temporary page images are removed after processing and are not retained by default.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (6)

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The PDF OCR workflow explicitly saves rendered page images to disk despite the documentation claiming the skill does not create or modify files. This inconsistency undermines trust boundaries and may expose sensitive document content through residual temporary files left on the system.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The skill is presented as 'read-only', but its documented behavior includes writing intermediate OCR image files and directing the agent to install packages. This mismatch can mislead users and policy layers about the skill’s actual side effects, increasing the chance of unintended filesystem changes or environment modification during use.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

Allowing Write conflicts with the skill’s declared read-only purpose and permits creation or modification of local files. Even if intended only for OCR intermediates, this creates an avoidable capability that could be abused for persistence, tampering, or data leakage staging.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

Allowing Write conflicts with the skill’s declared read-only purpose and permits creation or modification of local files. Even if intended only for OCR intermediates, this creates an avoidable capability that could be abused for persistence, tampering, or data leakage staging.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The OCR path sends image paths or URLs and potentially document-derived content to an external MCP service without a clear user-facing warning or consent step. For sensitive documents, this can result in unintended third-party disclosure of confidential data, which is a meaningful privacy and security risk.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
88% confidence
Finding

Natural-language instructions, triggers, parameters, workflow, and errors are all presented in Chinese, which effectively forces a specific language for users of the skill. There is no opt-in language choice or explanation that the skill is intended only for a Chinese-speaking or region-specific context.

Content

No source excerpt is available for this finding.