fuzzing-harness
Pass
Audited by Gen Agent Trust Hub on Oct 30, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill's primary purpose involves the ingestion and processing of untrusted data (binary bytes, crash inputs, and protocol packets).
- Ingestion points: The skill reads source code and external crash inputs to generate regression tests and harnesses.
- Boundary markers: The instructions explicitly mandate keeping execution local and contained, using containers or VMs with no network access, and stubbing external seams (database, network, etc.).
- Capability inventory: The skill uses
Read,Write, andGreptools. It also involves shell execution for building and running code, though these operations are not pre-approved and require explicit user consent. - Sanitization: The skill relies on environment isolation and user review rather than automated sanitization of the processed bytes.
- [EXTERNAL_DOWNLOADS]: The skill references and provides integration steps for several external tools and platforms.
- Evidence: References include Google's Atheris, ClusterFuzzLite, and OSS-Fuzz, as well as official documentation for LLVM, Rust-Fuzz, Go, and Trail of Bits.
- Reasoning: These references target trusted organizations and well-known, established security services. The downloads (e.g., GitHub Actions from Google's repository) are standard for the stated purpose of setting up CI/CD fuzzing.
Audit Metadata