The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's core functionality relies on ingesting and acting upon untrusted external data from research papers, competitor blogs, and leaderboard websites to drive its autonomous decision-making loop.
Ingestion points: The system monitors arxiv_queries, venues (e.g., ICLR, NeurIPS), and competitor_blogs as defined in Phase 3.1.
Capability inventory: The skill possesses the capability to write to the file system (Phase 5.1) and execute code/configurations during the experiment phase (Phase 5.2).
Sanitization: The instructions do not specify any validation or sanitization of the techniques extracted from external research before they are implemented as experiment hypotheses.
Boundary markers: There are no protective instructions or delimiters mentioned to ensure the agent ignores instructions embedded within the research data that might attempt to hijack the benchmarking process.
[DYNAMIC_EXECUTION]: The skill involves the automated generation and runtime execution of experimental logic (Phase 5). Because the inputs for these experiments are derived from external research summaries and architecture tweaks (Phase 3.2), there is a risk that the agent may generate and execute malicious code snippets if the external sources are compromised or maliciously crafted.
[EXTERNAL_DOWNLOADS]: The skill is designed to scrape various external domains to monitor the competitive landscape and ingest research. While academic sources like ArXiv are generally trustworthy, the inclusion of competitor blogs and arbitrary leaderboard URLs introduces untrusted external content into the agent's context.

auto-benchmark