The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to process data from external, untrusted sources such as cryptographic oracles and JWT headers, creating an attack surface for indirect prompt injection.
Ingestion points: Data ingested from target service responses in padding oracle scenarios (e.g., in cbc-padding-oracle.md and pkcs1-v1.5-bleichenbacher.md) and JWT payload components in jwt-alg-confusion.md.
Boundary markers: The instructions do not define explicit delimiters or security warnings to prevent the agent from interpreting instructions embedded within the processed cryptographic data.
Capability inventory: The skill encourages writing scripts that perform complex calculations and network requests, which could be influenced by malicious input.
Sanitization: There is no documentation regarding the sanitization or validation of the structure of responses received from external oracles.
[DATA_EXFILTRATION]: The skill describes techniques that require network operations to interact with external target services (oracles).
Evidence: Scenarios such as cbc-padding-oracle.md and ecb-prefix-oracle.md describe using tools like requests and curl to send crafted ciphertexts to a target URL and observe the results.
Context: These network operations are intended for target analysis rather than exfiltrating user credentials, but they constitute interactions with non-whitelisted external domains.
[EXTERNAL_DOWNLOADS]: The skill references and suggests the use of several external security libraries and research tools.
Evidence: Documentation suggests the use of fpylll, randcrack, yafu, gmp-ecm, and research repositories for specific attacks (e.g., mimoo/RSA-and-LLL-attacks, defund/coppersmith).
Context: The referenced resources are standard, well-known tools within the cryptography and security research community.

cryptography