Back to skill

Security audit

AB Agents Vision

Security checks for vulnerabilities and agentic risk

Overview

This image-analysis skill appears purpose-aligned, but it asks users to run mutable third-party code and executes an unpinned runtime package that can access their API key and submitted image data.

Review before installing. Use this only with images, screenshots, documents, and URLs you are comfortable sending to MiniMax, provide the API key deliberately, and avoid running the documented pipe-to-shell installer. Prefer a verified uv/uvx installation and a pinned, audited version of the MiniMax MCP package before use.

Vulnerability Patterns
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • 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 (3)

T03 ยท Remote Payload Retrieval and Execution

Error
Location
SKILL.md:27
Finding

Unverified Remote Installer Executed Directly by the Shell

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:27
Vulnerability Type: Remote payload retrieval and execution
Risk Level: High

Vulnerable Code:

bash
curl -LsSf https://astral.sh/uv/install.sh | sh

Technical Analysis

The installation instructions download a mutable script from an external URL and immediately pipe its contents into sh. The script is not pinned to a reviewed version, saved for inspection, or verified using a cryptographic signature or checksum.

Although astral.sh is the documented Astral domain, this construction makes the code executed by users dependent on the content returned at installation time. A compromise of the upstream website, deployment infrastructure, redirect destination, or signing account could replace the installer after the Skill has already passed review.

Installing software through an unrestricted remote shell script also exceeds the minimum privileges needed by the Skill's image-analysis functionality. The core functionality only requires an available, trusted client dependency; it does not inherently require executing mutable installation logic from the network.

Attack Path

  1. An attacker compromises the upstream installer, its hosting infrastructure, or a redirect destination followed because curl uses -L.
  2. The attacker modifies the response at https://astral.sh/uv/install.sh to contain arbitrary shell commands.
  3. A user follows the documented Quick Start instructions.
  4. curl retrieves the modified response and passes it directly to sh.
  5. The payload executes without an opportunity for inspection or integrity verification.

Impact Assessment

A substituted installer can execute arbitrary commands with all permissions held by the user running the installation. This can include reading or modifying user files, stealing credentials, installing persistence, downloading additional payloads, and accessing other resources available ...[truncated 242 chars]

Remediation
View remediation

Remediation Suggestions

  • Remove the curl | sh installation instruction.
  • Prefer a package-manager installation mechanism that provides versioning and integrity controls.
  • Alternatively, download a specific release artifact as a separate step, pin its version, and verify a publisher signature or documented SHA-256 checksum before execution.
  • Present download, verification, and installation as distinct commands so users can inspect the artifact.
  • Require uvx or an equivalent reviewed dependency to be installed before the Skill runs rather than installing it automatically with unrestricted shell code.
  • Explicitly advise users not to perform installation as root.

T08 ยท Insecure Dependencies

Error
Location
vision.sh:26
Finding

Unpinned Third-Party Package Downloaded and Executed at Runtime

Content
View full analysis

Vulnerability Details

File Location: vision.sh:26
Vulnerability Type: Insecure runtime dependency execution
Risk Level: High

Vulnerable Code:

bash
} | uvx minimax-coding-plan-mcp 2>/dev/null | python3 -c "

Technical Analysis

The script invokes minimax-coding-plan-mcp through uvx without an exact package version, lockfile, hash, or other provenance constraint. On an environment where the package is not already cached, uvx can retrieve the currently resolved package and execute its entry point.

Consequently, the effective executable can change independently of the reviewed Skill. A malicious new release, compromised publisher account, package-index compromise, or dependency-chain compromise could introduce arbitrary code without modifying either project file.

The process also inherits exported environment variables, including MINIMAX_API_KEY, and executes with the invoking user's operating-system permissions. Suppressing standard error with 2>/dev/null can additionally hide dependency-resolution warnings or runtime indicators from users.

Attack Path

  1. An attacker compromises the package publisher, package repository, or a transitive dependency and publishes malicious code under a version accepted by the unconstrained resolution.
  2. A user invokes vision.sh in an environment where uvx must resolve or download the package.
  3. uvx selects the attacker-controlled release because no audited version or integrity hash is pinned.
  4. The package entry point executes as the current user.
  5. The malicious package reads inherited credentials or accessible files and can transmit them over the network or modify the user's environment.

Impact Assessment

Successful exploitation provides arbitrary code execution with the permissions of the user invoking the Skill. The malicious dependency can access the exported MiniMax API key, inspect submitted image paths and prompts, read ot ...[truncated 373 chars]

Remediation
View remediation

Remediation Suggestions

  • Pin minimax-coding-plan-mcp to an exact, audited version.
  • Maintain a lockfile containing cryptographic hashes for the package and all transitive dependencies.
  • Install and verify dependencies as a separate deployment step rather than dynamically resolving them during every invocation.
  • Invoke a verified local executable after installation.
  • Review package provenance and monitor the pinned dependency for security updates.
  • Avoid suppressing all standard error output; preserve actionable resolution and verification failures.
  • Run the dependency in a sandbox with only the required environment variable, image access, and network destination where practical.

T09 ยท Insecure Skill Coding Practices

Warning
Location
vision.sh:25
Finding

User-Controlled Values Are Embedded into JSON-RPC Without Encoding

Content
View full analysis

Vulnerability Details

File Location: vision.sh:25
Vulnerability Type: JSON-RPC injection and malformed request construction
Risk Level: Medium

Vulnerable Code:

bash
{
    echo '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"vision","version":"1.0"}}}'
    sleep 1
    echo "{\"jsonrpc\":\"2.0\",\"id\":2,\"method\":\"tools/call\",\"params\":{\"name\":\"understand_image\",\"arguments\":{\"prompt\":\"$PROMPT\",\"image_source\":\"$IMAGE_PATH\"}}}"
} | uvx minimax-coding-plan-mcp 2>/dev/null | python3 -c "

Technical Analysis

PROMPT and IMAGE_PATH originate from command-line arguments and are interpolated directly into a JSON string. They are not encoded by a JSON serializer. Characters such as double quotes, backslashes, control characters, and embedded newlines can therefore terminate or alter their intended JSON string values.

A crafted argument can produce malformed JSON or inject additional fields or newline-delimited JSON-RPC messages into the MCP input stream. This is protocol-level injection rather than shell command injection: the variables remain inside a shell double-quoted argument, and the reviewed code does not demonstrate direct execution of their content as shell syntax.

Attack Path

  1. An attacker supplies or persuades a user to pass a crafted image path or prompt containing JSON delimiters and an embedded newline.
  2. Shell interpolation inserts the content into the echo argument without JSON escaping.
  3. The resulting stream contains a malformed request or attacker-controlled JSON-RPC structure.
  4. minimax-coding-plan-mcp parses any valid injected message or rejects the corrupted intended request.
  5. Depending on the MCP server's exposed methods and validation, the attacker may manipulate tool arguments, attempt unintended supported calls, or reliably cause the request to fail.

...[truncated 570 chars]

Remediation
View remediation

Remediation Suggestions

  • Construct every JSON-RPC message with a real JSON serializer rather than shell string interpolation.
  • Pass PROMPT and IMAGE_PATH to a small Python helper through argv or environment variables, then use json.dumps to encode the complete request.
  • Validate that image inputs use expected local-path or HTTPS URL formats.
  • Reject control characters where they are not needed.
  • Ensure the MCP server applies method allowlisting and schema validation independently of client-side encoding.
  • Add tests covering quotes, backslashes, Unicode, and newline characters in both arguments.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Tool MisuseTool Parameter Abuse, Chaining Abuse, Unsafe Defaults
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (8)

Chaining Abuse

High
Category
Tool Misuse
Confidence
98% confidence
Finding

The | sh pattern turns untrusted network content directly into shell commands, removing any opportunity for inspection and greatly increasing the blast radius of a compromised installer. In skill documentation, this is especially dangerous because users may copy-paste it verbatim, leading to immediate arbitrary command execution on their machine.

Content

Scanner excerpt ยท SKILL.md (reported line 27)May include surrounding context.

bash
# Install
curl -LsSf https://astral.sh/uv/install.sh | sh

# Set your MiniMax API key
export MINIMAX_API_KEY="sk-cp-your-key"

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The skill advertises local file and URL image analysis through the MiniMax VL API but does not clearly warn users that supplied images and extracted text will be transmitted to a third-party service. This can lead users to unknowingly send sensitive screenshots, documents, or internal URLs off-host, creating privacy, confidentiality, and compliance risk.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
86% confidence
Finding

The manifest describes a simple image-analysis wrapper for local files and URLs, but the code also searches ~/.minimax-env for MINIMAX_API_KEY. Reading credential material from a hidden file is not part of image understanding itself and expands the skill's effective access beyond the user-supplied image/prompt flow.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The script reads MINIMAX_API_KEY from the user's environment or from ~/.minimax-env, which is sensitive credential material. While it errors if the key is missing, there is no user-facing warning, comment, or other disclosure explaining that the skill accesses local credentials.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The skill sends the user-supplied image path or URL and prompt to a remote third-party API, but provides no explicit warning or consent boundary to the user. This is dangerous because users may pass sensitive local screenshots, documents, or internal URLs without realizing the contents will leave the host and be processed externally.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
93% confidence
Finding

The script executes an external MCP tool via uvx minimax-coding-plan-mcp without pinning an exact package version or digest. This creates a supply-chain risk: a malicious or compromised upstream release could be fetched and executed at runtime with the user's environment, including access to the API key and image data.

Content

No source excerpt is available for this finding.

External Script Fetching

Low
Category
Supply Chain
Confidence
95% confidence
Finding

The installation instructions fetch a remote script over the network and immediately execute it with the shell. If the remote host, transport, or delivered script is compromised, users can suffer arbitrary code execution during install.

Content

Scanner excerpt ยท SKILL.md (reported line 27)May include surrounding context.

bash
# Install
curl -LsSf https://astral.sh/uv/install.sh | sh

# Set your MiniMax API key
export MINIMAX_API_KEY="sk-cp-your-key"

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
84% confidence
Finding

The script invokes an external subprocess, uvx minimax-coding-plan-mcp, to perform the API interaction. Although subprocess use is central to the implementation, the script does not disclose this execution path to the user anywhere in its visible messaging or comments.

Content

No source excerpt is available for this finding.