The Agent Skills Directory

[METADATA_POISONING]: The skill uses documentation links pointing to the domain sosumi.ai instead of the official Apple Developer documentation at developer.apple.com. This is deceptive content that redirects users to a third-party controlled domain for security-critical implementation details.
Evidence in SKILL.md: References to https://sosumi.ai/documentation/devicecheck/dcdevice and https://sosumi.ai/documentation/devicecheck/dcappattestservice.
Evidence in references/device-integrity-patterns.md: References to https://sosumi.ai/documentation/devicecheck/validating-apps-that-connect-to-your-server and https://sosumi.ai/documentation/devicecheck/assessing-fraud-risk.
[INDIRECT_PROMPT_INJECTION]: The skill demonstrates patterns for ingesting data from external server endpoints, which creates a surface for processing untrusted data within the agent's execution flow.
Ingestion points: The fetchServerChallenge method in SKILL.md retrieves data from a network endpoint and injects it into the clientDataHash calculation.
Capability inventory: The skill has the capability to perform network operations (URLSession), manage hardware-backed keys in the Secure Enclave, and read/write to the system Keychain.
Boundary markers: The implementation lacks boundary markers or explicit instructions for the agent to ignore potentially malicious content embedded in the server responses.
Sanitization: While the challenge data is hashed using SHA-256 for the App Attest flow, there is no validation or sanitization of the server response content before it is processed by the agent's logic.

device-integrity