The Agent Skills Directory

[DATA_EXFILTRATION]: Step 6 in SKILL.md instructs the agent to upload UI screenshots to catbox.moe, a public file hosting service. While the skill warns to store these as temporary local artifacts and verify links, catbox.moe provides permanent public access to uploaded files, which can lead to the accidental exposure of sensitive UI information or PII.
[COMMAND_EXECUTION]: The skill makes extensive use of external command-line tools including git, gh, trufflehog, and caffeinate, as well as several AI review engines (codex, opencode, coderabbit). The scripts autoreview_test.py and test-review-harness.py demonstrate the use of subprocess.run to execute these binaries.
[CREDENTIALS_UNSAFE]: The file tests/fixtures/typescript-sensitive-literals.ts contains several hardcoded fake credentials, including OpenAI API keys (sk-proj-...), GitHub tokens (ghp_...), and AWS Access Keys (AKIA...). While these are explicitly marked as "FAKE" and intended for testing the secret scanner logic, they remain hardcoded string literals resembling real secrets.
[INDIRECT_PROMPT_INJECTION]: Step 5 in SKILL.md implements a triage ledger system where findings previously marked as wont-fix or accepted-risk are stored in /tmp/send-it-findings--.md and subsequently interpolated into the prompt for the AI review engines using the --prompt flag. This creates a surface for indirect prompt injection if the content of the ledger (derived from tool outputs or user triage) is maliciously crafted.
[DYNAMIC_EXECUTION]: The test script scripts/autoreview_test.py uses importlib.util and SourceFileLoader to dynamically load and execute the autoreview script from the filesystem for testing purposes.
[EXTERNAL_DOWNLOADS]: The skill uses curl to upload files to catbox.moe and requires the installation of several external dependencies (git, gh, trufflehog, coderabbit) to function, which may involve downloading code or binaries from remote servers.

send-it