Uncategorized

add-educational-comments

Gold
1 Active Threats Found

Click to view details and understand why these patterns are dangerous.

5,856 tokensgemini-3-pro-preview

Safety Score

75/ 100

Detected Capabilities

shell

Sensitive Files

    Clean filesystem scan

Deep Audit Findings

The skill instructs an AI agent to read, modify, and rewrite code files to add educational comments, while also allowing it to fetch external URLs for references. The primary security concerns include arbitrary file modification (path traversal) and potential Server-Side Request Forgery (SSRF) via the user-controlled 'Fetch List'. Additionally, the static analyzer flagged a potential injection vector due to unusual instructions simulating user keyboard input. Proper sandboxing and human-in-the-loop validation are necessary.

Arbitrary File Modification and Path Traversal Risk

The agent is instructed to modify a 'target file(s)' specified by the user. Without proper path validation or sandboxing controls at the framework level, an attacker could instruct the agent to read or overwrite sensitive system files, configuration files, or scripts outside the intended workspace.

MEDIUM
SKILL.md
Add educational comments to the file specified, or prompt asking for file to comment if one is not provided.

Potential Server-Side Request Forgery (SSRF) via Fetch List

The prompt supports an optional `Fetch List` for URLs. If the agent's underlying runtime automatically fetches these URLs without validation, a malicious user could provide local IP addresses, cloud metadata URLs, or internal endpoints, turning the agent into a proxy for internal network scanning.

LOW
SKILL.md
- **Fetch List**: Optional URLs for authoritative references.

Suspicious Keystroke/Input Instruction

The prompt contains an unusual instruction to 'Input data as if typed on the user's keyboard.' While likely intended to enforce plain text output without copy-paste artifacts, in certain automation frameworks, this could be misinterpreted as a command to simulate keystrokes, potentially leading to unintended OS-level interactions.

LOW
SKILL.md
Input data as if typed on the user's keyboard.

Risk of Code Corruption via Automated Edits

The agent is instructed to increase the file line count by 125%. LLMs are prone to hallucinations, and forcing such a massive output volume into existing code could inadvertently break execution, alter namespaces, or introduce subtle logic bugs despite the prompt's warnings.

LOW
SKILL.md
Increase the total line count by **125%** using educational comments only (up to 400 new lines).

Attack Surface Chain

1

An attacker provides a malicious target file path (e.g., `../.env` or `/etc/hosts`) or includes an internal IP in the 'Fetch List' parameter.

2

The agent executes its prompt instructions, attempting to read the targeted out-of-bounds file or making an HTTP request to the internal URL.

3

The agent rewrites the targeted sensitive file with comments, causing system malfunction or data leakage, or returns information from the internal network (SSRF).