The Agent Skills Directory

[COMMAND_EXECUTION]: The skill leverages the Bash tool to execute localized Python scripts for deterministic citation verification against academic databases.
Evidence: agents/bibliography_agent.md contains instructions for the agent to run uv run python skills/core/alterlab-citation-verifier/scripts/verify_citations.py to check the existence of curated sources.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a vulnerability surface for indirect prompt injection due to its core function of ingesting and synthesizing external, untrusted academic literature and web search results.
Ingestion points: Untrusted data enters the context via the WebSearch and WebFetch tools used by the bibliography_agent.md and source_verification_agent.md during literature searches and fact-checking.
Boundary markers: While the skill provides structured output templates and checkpoints for human review (e.g., in SKILL.md and references/orchestration_workflows.md), it lacks explicit technical delimiters or specific "ignore embedded instructions" warnings for the raw data ingestion phase.
Capability inventory: The agent team has broad capabilities, including Read, Write, Edit, Bash, WebFetch, and WebSearch, which could be exploited if malicious instructions in external papers are followed.
Sanitization: No specific technical sanitization or filtering of external content is described beyond the agent's internal evaluation logic and the use of the alterlab-citation-verifier script.

alterlab-deep-research