The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill references several external source code repositories for security tooling:
Fetches the AFL++ source code from its official GitHub repository: https://github.com/AFLplusplus/AFLplusplus.
Downloads the honggfuzz fuzzer from Google's official GitHub organization: https://github.com/google/honggfuzz.
Obtains the afl-cov coverage tool from its GitHub repository: https://github.com/mrash/afl-cov.
[PRIVILEGE_ESCALATION]: The instructions include the use of sudo for administrative operations, which is consistent with standard system maintenance:
Installing packages via system managers (apt, dnf, pacman).
Installing compiled software to system directories using make install.
[COMMAND_EXECUTION]: The skill involves the execution of various commands to facilitate software testing:
Compilation of target programs and harnesses using clang, gcc, and make.
Running the afl-fuzz and libFuzzer engines to dynamically test binaries.
The skill explicitly emphasizes the security principle of running all dynamic testing inside a sandbox environment.
[INDIRECT_PROMPT_INJECTION]: The skill defines a surface for processing external data during the fuzzing process:
Ingestion points: Untrusted data enters the context through seed files in the in/ directory and seed_sample.bin (SKILL.md).
Boundary markers: The instructions mandate the use of a sandbox ([[re-sandbox]]) to isolate dynamic execution and mitigate potential risks from processing malicious inputs.
Capability inventory: The workflow utilizes powerful system capabilities including compilers, build tools, and the afl-fuzz engine (SKILL.md).
Sanitization: While no specific input sanitization is described, the workflow relies on the target's internal checks and host sandboxing to manage the risks associated with processing adversarial inputs.

re-fuzzing