The Agent Skills Directory

[DATA_EXFILTRATION]: The skill packages and uploads manuscript directories and source code archives to an external public file-sharing service (catbox.moe) via curl. While secret scrubbing routines are implemented, utilizing an anonymous public hosting platform for project logs and source code creates a risk of exposing sensitive or proprietary data.
[INDIRECT_PROMPT_INJECTION]: The skill ingests history logs and conversational data to identify defects, presenting an indirect prompt injection vulnerability surface.
Ingestion points: Ingests conversation transcripts from ~/.claude/projects/<project-slug>/*.jsonl and build/smoke logs from $RUN_DIR/proofs/ in phases/01-gather-evidence.md and phases/02-mine-the-session.md.
Boundary markers: Lacks explicit boundary delimiters or system instructions to encapsulate historical log text as untrusted data.
Capability inventory: Possesses capabilities to execute arbitrary local shell commands via the Bash tool, interact with GitHub repositories via the gh CLI, and post files via curl.
Sanitization: Contains regex-based scanning for credentials and basic PII shapes (references/secret-scrubbing.md), but lacks validation or structural escaping to neutralize embedded natural language instructions before processing.

printing-press-retro