The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The protocol is designed to handle data from external and untrusted sources such as web searches, public filings, and registries, creating a surface for potential indirect prompt injection attacks.
Ingestion points: The agent is instructed to gather intelligence from the 'world' (public sources, search engines, and trade data) as part of the search-plan gate.
Boundary markers: The protocol uses a mandatory search-plan gate and stable output schemas to maintain structural control, though it does not define specific technical delimiters for sanitizing raw input text.
Capability inventory: The skill is entirely markdown-based with no scripts, binaries, or automated tool invocations, which restricts the potential impact of any malicious payload embedded in research data.
Sanitization: The framework enforces high-integrity evidence labeling (Fact, Inference, Assumption) and a 'confidence stacking' rule to verify signals across multiple independent channels, effectively reducing the risk of acting on poisoned data.
[NO_CODE]: The skill consists exclusively of markdown instructions, templates, and examples. It does not include any executable scripts, configuration files for package managers, or binary assets.

autonomous-investigation