The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes Python scripts and C source files intended for local execution and compilation during development and evaluation. These include demonstration fuzzers (e.g., coverage_fuzzer.py) and C harnesses (e.g., LLVMFuzzer_harness.c) that utilize standard development tools like gcc and python. These operations are limited to the local environment and are central to the skill's stated purpose.
[INDIRECT_PROMPT_INJECTION]: The skill addresses the vulnerability surface where fuzzers process untrusted input data. It provides mandatory evidence chains and specific remediation guidance, such as implementing strict size guards (if (size < 3)) and ensuring statelessness through per-input resets (memset(&st, 0, sizeof(st))), which are best practices for securing fuzzing harnesses.
[SAFE]: Several files (e.g., LLVMFuzzer_harness_bad.c and naive_fuzzer.py) contain intentionally incorrect logic, such as out-of-bounds reads or lack of state resets. These are clearly labeled as 'BAD' and 'intentionally incorrect', serving as educational fixtures for synthetic evaluation to ensure the agent can distinguish between a correct and an incorrect harness.

fuzzing-harness-kernel