The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The file references/cli-actorization.md includes a Dockerfile template that executes a remote script from an external, non-trusted source using curl | sh. This specifically targets https://raw.githubusercontent.com/houseabsolute/ubi/master/bootstrap/bootstrap-ubi.sh. This practice bypasses integrity checks and is explicitly warned against in the main SKILL.md file, creating a significant security contradiction within the skill's own instructions.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to build actors that process data from external websites. This introduces a vulnerability surface where untrusted data (crawled HTML, text, or metadata) is ingested into the actor's context.
Ingestion points: Data is ingested via Actor.getInput() (JS), Actor.get_input() (Python), or apify actor:get-input (CLI), as well as through direct crawling of external web content.
Boundary markers: The skill provides a 'Security' section in SKILL.md advising on sanitization and validation, but these rely on correct implementation by the developer and do not provide automated boundaries.
Capability inventory: Actors have capabilities including file writing (Datasets/KV stores), network operations (Apify API), and command execution via apify run and apify push.
Sanitization: The skill provides high-level guidance for sanitization but lacks specific, enforced sanitization logic in its templates.
[EXTERNAL_DOWNLOADS]: The skill references several external resources, including scripts from the houseabsolute GitHub repository and configuration templates from the apify/actor-templates repository. While the Apify resources are considered vendor-internal and safe, the third-party script from houseabsolute is not from a recognized trusted source.
[COMMAND_EXECUTION]: The skill makes extensive use of CLI commands (apify init, apify run, apify push) and environment variables (APIFY_TOKEN). It also includes shell script templates in references/cli-actorization.md that execute commands based on input parsed from JSON using jq. While the provided example uses variable quoting ("$MY_PARAM"), the pattern of passing external input into shell commands remains a potential risk factor if not strictly validated.

apify-actorization