The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill requires cloning source code from external GitHub repositories (FasterDecoding/Medusa and hao-ai-lab/LookaheadDecoding) that are not associated with trusted organizations or well-known services defined in the security guidelines.\n- [REMOTE_CODE_EXECUTION]: The installation instructions for the Medusa and Lookahead Decoding tools involve running pip install -e . on the cloned repositories. This is a high-risk operation as it executes the setup.py or pyproject.toml files from these untrusted external sources, allowing for arbitrary code execution on the user's system during the installation process.\n- [COMMAND_EXECUTION]: The documentation provides multiple shell commands (git clone, pip install, cd) for setting up the environment. These commands are intended to be executed in a shell environment, which can be exploited if the source repositories are compromised.\n- [PROMPT_INJECTION]: The skill established a surface for indirect prompt injection by processing external data as LLM prompts.\n
Ingestion points: Untrusted user input enters the agent context via the prompt variable in the code snippets provided in SKILL.md.\n
Boundary markers: There are no delimiters or 'ignore embedded instructions' warnings used when interpolating the prompt variable into the model generation calls.\n
Capability inventory: The skill environment (as described in the installation steps) includes the capability to execute shell commands and install packages, which could be abused if an injected prompt misleads the agent.\n
Sanitization: The provided Python code lacks any validation or filtering of the content within the prompt variable before it is processed by the model.

speculative-decoding