The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill instructs users to manually login to various platforms and store sensitive session cookies in the ~/.playwright-data/ directory. These credentials are then accessed by the publisher.py and index_checker.py scripts to perform automated actions.
[COMMAND_EXECUTION]: The skill utilizes the Playwright library to automate web browser interactions, enabling the agent to navigate websites, fill forms, and click buttons programmatically.
[EXTERNAL_DOWNLOADS]: The skill fetches research data from external search providers like Baidu and Bing, and interacts with third-party AI chat interfaces (Doubao, Qianwen, DeepSeek). It also requires downloading browser binaries via playwright install chromium.
[DATA_EXFILTRATION]: The automated publishing mechanism (publisher.py) moves content from the local environment to external platforms. While intended for article posting, this functionality could be repurposed for unauthorized data transmission if the agent's instructions are compromised.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it ingests untrusted content from the web and uses it to build prompts for further LLM tasks.
Ingestion points: Search result snippets and AI-generated responses fetched in scripts/competitor_research.py and scripts/index_checker.py.
Boundary markers: Prompts in scripts/article_generator.py interpolate external data without specific delimiters or instructions to ignore embedded commands.
Capability inventory: File system write access to the data/ directory and network operations via httpx and playwright browser automation.
Sanitization: There is no evidence of filtering or escaping performed on the external research data before it is presented back to the LLM context.

geo-agent