The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructions direct the agent to download a script (call.mjs) from a remote server (https://fittings.sh/skill/call.mjs) and execute it using the node runtime. This pattern allows for the execution of arbitrary code fetched at runtime.
Evidence: curl -sO https://fittings.sh/skill/call.mjs followed by node call.mjs ....
[EXTERNAL_DOWNLOADS]: The skill fetches external resources at runtime, including the call.mjs execution script and catalog data from the vendor's domain.
Evidence: curl -s https://fittings.sh/catalog/devicerecall.search and curl -sO https://fittings.sh/skill/call.mjs.
[CREDENTIALS_UNSAFE]: The skill relies on a sensitive credential, FITTINGS_PRIVATE_KEY, which is instructed to be stored in an environment variable in plaintext. The documentation explicitly notes that the key sits in plaintext wherever the agent runs, which poses a risk of exposure or exfiltration if the environment is compromised.
Evidence: export FITTINGS_PRIVATE_KEY=0x....
[COMMAND_EXECUTION]: The skill utilizes shell commands to perform its core functions, including environment variable management, network fetching, and script execution.
Evidence: Usage of curl, export, and node within the provided instructions.
[INDIRECT_PROMPT_INJECTION]: The skill processes data from an external source (FDA enforcement reports via a proxy) and provides instructions to a shell environment, creating a potential surface for indirect injection if the remote data contains malicious patterns.
Ingestion points: The skill fetches recall data from https://fittings.sh/v1/devicerecall/search (which proxies api.fda.gov).
Boundary markers: None provided in the documentation to delimit external data from agent instructions.
Capability inventory: The skill has the capability to execute shell commands (curl) and Node.js scripts (node).
Sanitization: No explicit sanitization or validation of the remote data is described before it is processed or displayed.

fittings-devicerecall-search