The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the agent to clone external repositories from github.com/GPTomics/bioSkills and github.com/ClawBio/ClawBio and execute scripts or pipelines contained within them. This allows for the execution of untrusted code from sources that are not recognized as trusted vendors.
[PRIVILEGE_ESCALATION]: The instructions command the agent to use sudo apt install for installing system-level bioinformatics tools. Invoking sudo grants the agent elevated permissions and risks full system compromise.
[EXTERNAL_DOWNLOADS]: The skill performs shallow clones of external GitHub repositories to the /tmp directory. These downloads originate from entities outside the known trusted infrastructure.
[COMMAND_EXECUTION]: The skill executes multiple shell commands to modify the system environment, including pip install, Rscript -e 'BiocManager::install()', and conda install, which introduce unverified third-party dependencies.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to fetch and read documentation and instructions from external repositories (e.g., using cat on SKILL.md or README.md) to guide its behavior. It lacks any boundary markers or sanitization to prevent malicious instructions embedded in that content from influencing the agent.
Ingestion points: cat commands reading from cloned repository files in /tmp.
Boundary markers: None present to distinguish external content from trusted instructions.
Capability inventory: Includes git clone, pip install, sudo apt install, and execution of external shell scripts.
Sanitization: No filtering or validation of the fetched content is performed.

bioinformatics