Back to skill

Security audit

Gemini Watermark

Security checks for vulnerabilities and agentic risk

Overview

This skill is transparent and local, but it is built to remove visible AI provenance marks and encourages sharing cleaned images without warning users about authorization or platform-policy risks.

Review this before installing if you plan to publish or share outputs. Only remove watermarks when you have clear rights and doing so does not violate platform terms, licensing terms, or disclosure expectations. Install dependencies in an isolated environment and prefer pinned, reviewed versions.

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
SKILL.md:22
Finding
Unpinned Third-Party Dependency Installation## Vulnerability Details **File Location**: `SKILL.md`, lines 22–25 **Vulnerability Type**: Unpinned dependencies and mutable supply-chain inputs **Risk Level**: Medium ```bash pip install Pillow numpy # Recommended: use uv for faster, isolated installs uv pip install Pillow numpy ``` ### Technical Analysis The installation instructions retrieve `Pillow` and `numpy` without version constraints, artifact hashes, a lockfile, or an explicitly trusted package index. Consequently, package resolution depends on mutable repository state and the user's package-manager configuration. A future compromised release, malicious package-index configuration, or tampered distribution artifact could cause unreviewed installation logic or native code to execute. The instructions also conflict with the documentation's claims that there are “no downloads” and “no compiled binaries”: pip and uv ordinarily download package artifacts, and both dependencies may be distributed as platform-specific binary wheels. The package names are correctly spelled, and the reviewed files contain no evidence that the current dependency releases or configured sources are malicious. The risk arises from accepting unpinned and unverifiable future supply-chain inputs. ### Attack Path 1. A user follows the documented setup instructions. 2. pip or uv connects to the package index configured in the user's environment. 3. The resolver selects the latest compatible `Pillow` and `numpy` releases rather than versions reviewed with this Skill. 4. If a selected artifact, upstream release, mirror, or configured index has been compromised, the package manager downloads and installs attacker-controlled content. 5. Installation-time logic or imported dependency code executes with the privileges of the user running the installation or Skill. ### Impact Assessment Successful exploitation could execute arbitrary code with the privileges of the user who installs or runs th ...[truncated 430 chars]
Remediation
## Remediation Suggestions 1. Pin dependency versions that have been reviewed and tested, preferably through a committed lockfile or requirements file. 2. Pin exact artifacts with cryptographic hashes and install them using hash verification, for example: ```bash python3 -m pip install --require-hashes -r requirements.txt ``` 3. Document and enforce the expected trusted package index rather than implicitly accepting arbitrary user-configured mirrors. 4. Use an isolated virtual environment and avoid installing dependencies with administrator or root privileges. 5. Add automated dependency vulnerability and integrity scanning to the release process. 6. Update the documentation to state accurately that dependencies may be downloaded and may include native binary wheels.
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
Findings (1)

Missing User Warnings

Medium
Confidence
96% confidence
Finding
This is a true issue. The skill is explicitly designed to remove visible Gemini watermarks and even encourages users to 'clean images before publishing or sharing' without any warning about legal, contractual, provenance, or platform-policy risks. In this context, omission of safeguards materially increases misuse potential by facilitating deceptive redistribution of AI-generated content after removal of attribution or authenticity signals.

Static analysis

No suspicious patterns detected.