The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its processing of untrusted web search results. The resolve_markdown_links.py script extracts link titles and URLs from a report file and prints them to the console. Because the agent is instructed to act on this output (e.g., re-searching for homepage links or fixing broken ones), an attacker could craft a malicious web page title that, when reflected in the script output, provides unauthorized instructions to the agent. Ingestion points: External data from web and video searches is compiled into a markdown file. Boundary markers: No explicit markers or instructions to ignore embedded commands are present in the synthesis workflow. Capability inventory: The skill executes local Python scripts and makes network requests. Sanitization: The script does not sanitize the extracted link text before displaying it back to the agent.
[COMMAND_EXECUTION]: The skill requires the agent to execute shell commands for initial environment discovery (e.g., uname -a, echo $SHELL) and to run the included Python utility script (resolve_markdown_links.py) for link verification.
[DATA_EXFILTRATION]: The skill makes network requests to verify the status of arbitrary external URLs found during research. While these status checks (HEAD/GET requests) are necessary for link validation, they involve outgoing connections to non-whitelisted domains.

synthesizing-knowledge