T09 · Insecure Skill Coding Practices
- Location
scripts/generate_submit_run_body.sh:77- Finding
Arbitrary Python Code Execution Through Unsafe Input Interpolation
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill is for AI image/video generation, but it asks users to provide browser login cookies, stores them in plaintext, and includes a helper script vulnerable to local code execution from crafted input.
Review before installing. This skill may act as your logged-in Pippit or Xiaoyunque account, consume account resources, and expose session cookies if chat logs, the workspace, or the config file are accessible. It should use a scoped auth flow or secure secret store and fix the helper script before normal use; avoid pasting browser session cookies into chat.
scripts/generate_submit_run_body.sh:77Arbitrary Python Code Execution Through Unsafe Input Interpolation
SKILL.md:42Browser Session Cookies Are Collected Through Chat and Persisted in Plaintext
This is a direct request for users to extract browser session cookies and hand them to the agent, followed by instructions to store them for later reuse. Because these cookies authenticate as the user, disclosure or misuse can grant full account access, enable impersonation, and expose user data or paid resources; the broad skill trigger and file persistence make the danger higher, not lower.
The skill explicitly instructs the agent to collect browser-derived session cookies and persist them in a local config file for reuse. Session cookies are bearer credentials; storing them locally without strong safeguards or user warning risks account takeover if the file, logs, or workspace are exposed.
The trigger scope is broad enough that the skill may activate on generic image/video-related requests without clear user intent to use this specific service. That increases the chance the agent will invoke workflows that request or use sensitive session cookies for the wrong task or wrong service context.
These instructions direct the agent to automatically infer whether the user should be served in the domestic or overseas region and to keep that choice stable across the conversation. Because region selection can determine locale and service behavior, this creates an implicit locale policy decision without explicit user choice except as a fallback when inference fails.
The skill directs the agent to create and modify a config file containing persistent credentials and account identifiers, but provides no protections around secret storage, retention, masking, or access control. This creates a durable secret persistence mechanism that can be abused by other tools, future prompts, or anyone with filesystem access.
The script loads uid, workspace_id, and space_id from a local config file and unconditionally embeds them into the generated JSON output. If this output is logged, copied into tickets, shared in chats, or used in less-trusted environments, those identifiers may be exposed and enable account or workspace enumeration, unintended API use, or correlation of internal resources.