Back to skill

Security audit

ADME Property Predictor

Security checks for vulnerabilities and agentic risk

Overview

The skill is a local drug-property prediction helper with some sloppy documentation and unpinned dependencies, but no hidden data theft, persistence, or destructive behavior was found.

Before installing, run it in an isolated environment, pin or review the RDKit dependency, remove the unnecessary dataclasses dependency for Python 3.10+, and provide only molecule SMILES or trusted CSV input files. Treat results as screening estimates, not experimental or clinical evidence, and use explicit output paths because the script can overwrite files. Until fixed, do not rely on --properties to limit calculation scope.

Vulnerability Patterns
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • 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
Findings (1)

T08 · Insecure Dependencies

Warning
Location
requirements.txt:1
Finding

Unpinned Third-Party Dependency

Content
View full analysis

Vulnerability Details

File Location: requirements.txt:1-2
Vulnerability Type: Unpinned third-party dependencies
Risk Level: Medium

Vulnerable Code

text
dataclasses
rdkit

Technical Analysis

The project declares third-party dependencies without exact versions or verified integrity hashes. Consequently, separate installations may resolve to different package releases without any corresponding repository review. This prevents reproducible builds and exposes installation to upstream package compromise, unexpected releases, or dependency-resolution changes.

The dataclasses dependency is also unnecessary for the documented Python 3.10+ baseline because dataclasses is included in the Python standard library. Retaining an unnecessary external package increases the supply-chain attack surface.

The audit found no evidence that either currently declared package is malicious. The vulnerability is the unconstrained and unverifiable dependency-resolution process.

Attack Path

  1. A user or deployment system runs a package installation command using requirements.txt.
  2. The package resolver selects the package versions available from the configured registry at installation time.
  3. If an upstream account, release, registry, or resolution path has been compromised, the resolver downloads the affected artifact.
  4. Malicious installation or import-time code executes with the privileges of the user or service performing the installation.
  5. Because the project specifies neither trusted hashes nor reviewed versions, the installation process does not detect that the resolved artifact differs from the dependency version previously reviewed.

Impact Assessment

Successful exploitation could execute arbitrary code with the privileges of the account installing or running the Skill. Depending on its environment, this may permit access to project files, process environment variables, user-readable credentials, generated ADME results, and ...[truncated 321 chars]

Remediation
View remediation

Remediation Suggestions

  1. Remove dataclasses for the documented Python 3.10+ runtime because it is provided by the standard library.
  2. Pin RDKit to an explicitly reviewed version:
    text
    rdkit==<reviewed-version>
    
  3. Generate and commit a lock file containing hashes for all direct and transitive dependencies.
  4. Enforce hash verification during installation, such as with pip install --require-hashes.
  5. Install only from an approved package index over authenticated TLS, and disable untrusted extra indexes.
  6. Use automated dependency scanning and a controlled update process to review new releases before changing the lock file.
  7. Build and run the Skill in a least-privileged, isolated environment with access only to necessary input and output paths.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • 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
Findings (11)

Tp4

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

The skill description promises a bounded, validated, reproducible analysis workflow, but the observed behavior includes concrete file I/O and ADME prediction logic that is only partially reflected in the contract. This mismatch can cause operators or automated routers to trust the skill for safer, narrower behavior than it actually performs, leading to unsafe invocation or mishandling of user data.

Content

No source excerpt is available for this finding.

Ae4

Medium
Category
analysis-evasion
Confidence
80% confidence
Finding

Suspicious Unicode normalization or mixed-script content

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
81% confidence
Finding

The skill exposes file read/write behavior through its packaged workflow but does not declare any tool scope or permissions boundary. In agent environments, this weakens policy enforcement and can let the skill be invoked with broader filesystem access than reviewers or orchestrators expect.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The activation guidance is broad enough to match generic data-analysis tasks beyond molecular ADME prediction. Over-broad routing criteria are dangerous in agent ecosystems because they increase the chance that the skill is auto-selected for inappropriate requests, leading to erroneous processing, file access, or misleading outputs outside its competence.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The example invocation uses free-text clinical-style input rather than molecular input, directly conflicting with the documented ADME purpose. In an agent setting, contradictory examples can trigger incorrect routing or make downstream systems pass arbitrary text into chemistry tooling, increasing the chance of failures, misprocessing, or accidental misuse.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
87% confidence
Finding

The generic required-input guidance conflicts with the concrete parameter contract, obscuring what inputs are actually mandatory for safe execution. This can lead agents or users to supply ambiguous data, causing the skill to operate on unintended files or enter inconsistent fallback paths that are not adequately constrained.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The manifest describes a reproducible, explicitly validated workflow with structured outputs, and the CLI reinforces this by offering a --properties selector for limited analysis. However, predict() accepts properties and then ignores it entirely, calculating absorption, distribution, metabolism, excretion, and overall recommendation for every run, so the actual behavior exceeds the apparent scoped-analysis intent presented by the interface and documentation.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The inline CLI documentation explicitly states --properties is for selecting 'Specific properties to calculate'. In practice, the value is passed into predict() at L473 but never used there, which directly contradicts the command help and creates an intent-code divergence for users relying on scoped execution.

Content

No source excerpt is available for this finding.

Unpinned Dependencies

Low
Category
Supply Chain
Confidence
91% confidence
Finding

The dependency is unpinned, so installs may resolve to different versions over time, reducing reproducibility and increasing supply-chain risk if a future release introduces a vulnerability or malicious code. In a workflow-oriented skill that emphasizes reproducibility, this is a real security and integrity weakness even though the file itself is not overtly malicious.

Content

Scanner excerpt · requirements.txt (reported line 1)May include surrounding context.

text
dataclasses
rdkit

Unpinned Dependencies

Low
Category
Supply Chain
Confidence
95% confidence
Finding

The unpinned rdkit dependency can cause non-reproducible environments and exposes consumers to supply-chain risk from unexpected upstream releases. Because rdkit is a substantial third-party package used for scientific computation, an unreviewed version change could affect both security posture and analytical correctness in this skill's domain.

Content

Scanner excerpt · requirements.txt (reported line 2)May include surrounding context.

text
dataclasses
rdkit

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
84% confidence
Finding

This code writes prediction results to a user-specified path, but the only disclosure appears after the write succeeds. There is no prior confirmation prompt, warning comment/docstring, or explicit notice that running with --output will create or overwrite a file.

Content

No source excerpt is available for this finding.