The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines a framework for building validation logic that ingests tool execution parameters, which can be manipulated by external inputs to attempt to bypass the intended security constraints.
Ingestion points: Data enters through the params and normalisedParams arguments in the shouldBlockPreToolExecution and shouldBlockPostParamsNormalization methods, as seen in SKILL.md and examples/max-amount-policy/index.ts.
Boundary markers: The instructions recommend filtering by tool method names but do not explicitly instruct developers to use delimiters or instruction-isolation techniques when parsing parameters.
Capability inventory: The policies are intended to guard high-risk operations including transfer_hbar, transfer_fungible_token, and delete_account as listed in SKILL.md.
Sanitization: The provided example in examples/max-amount-policy/index.ts demonstrates basic input validation using Number() and Number.isFinite() for numerical values, but does not cover general-purpose sanitization for complex objects.

Hedera Policy Creation