The Agent Skills Directory

[PROMPT_INJECTION]: The documentation file references/sdks/integrations/superagent.md contains the explicit prompt injection payload "ignore all previous instructions" inside a code example. While intended for illustrative purposes, if an AI agent reads this documentation as part of its instructions, it may interpret the string as a direct command to bypass its current constraints.
[REMOTE_CODE_EXECUTION]: Multiple documentation files (references/sdks/cli.md, references/sdks/integrations/codex.md, references/sdks/integrations/skills.md) instruct users and agents to install software using a high-risk piped execution pattern: curl -fsSL https://hindsight.vectorize.io/get-cli | bash. Piped remote execution is a critical attack vector for arbitrary command execution.
[EXTERNAL_DOWNLOADS]: The skill references first-party installation scripts and documentation from domains such as hindsight.vectorize.io and hindsight.dev. Neither of these domains is listed in the authorized naming patterns for the skill's author ('vectorize-io'), nor do they appear on the trusted organizations list. This discrepancy poses a typosquatting risk, as vectorize-io.io is listed as a legitimate pattern while vectorize.io is not.
[INDIRECT_PROMPT_INJECTION]: The documentation files act as an ingestion surface for untrusted data. The lack of strict boundary markers or sanitization for documentation reading tasks creates a vulnerability where malicious instructions embedded in the doc text (like those found in the Superagent integration example) can influence agent actions.
[NO_CODE]: The skill consists entirely of documentation in Markdown and structured configuration in JSON format; it does not contain any executable script files (.py, .js, .sh) within the provided folder structure.

hindsight-docs