The Agent Skills Directory

Prompt Injection (SAFE): No direct prompt injection, jailbreak attempts, or override markers were detected in the instructions.
Data Exposure & Exfiltration (SAFE): No sensitive file paths (e.g., SSH keys, AWS credentials) are accessed, and no unauthorized network exfiltration patterns were identified.
Obfuscation (SAFE): No multi-layer Base64, zero-width characters, or homoglyphs are present. Hexadecimal escapes in the dictionary examples are standard for describing binary protocols and do not hide malicious intent.
Unverifiable Dependencies (LOW): The skill suggests installing 'afl++' via 'apt'. While this involves package installation, 'apt' is a trusted system package manager and AFL++ is a well-known security tool.
Indirect Prompt Injection (LOW): The skill creates an attack surface by instructions that process external, potentially untrusted data.
Ingestion points: fuzz_parser.c, myparser.c, and files within the corpus/ directory.
Boundary markers: Absent; the skill does not instruct the agent to ignore instructions embedded within the source code or corpus files it processes.
Capability inventory: Compilation and execution via clang, execution of fuzzer binaries (./fuzz_parser), and system tool usage (afl-fuzz).
Sanitization: Absent; the workflow assumes the user provides the files to be fuzzed.
Dynamic Execution (LOW): The skill involves compiling and running C/C++ code at runtime. This is the primary and intended purpose of a fuzzing skill; however, it remains a capability that an agent must handle with caution. Severity is downgraded from MEDIUM to LOW per [TRUST-SCOPE-RULE] as it is essential to the skill's utility.

fuzzing