ai-fuzzing
Pass
Audited by Gen Agent Trust Hub on Sep 19, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [COMMAND_EXECUTION]: The skill extensively uses the
Bashtool to execute fuzzing engines, instrument code, and process crash data. This behavior is fundamental to the skill's purpose of automating vulnerability discovery. - [EXTERNAL_DOWNLOADS]: The skill provides instructions to download and install various security tools and grammars from external repositories (e.g., AFL++, Nautilus, Gramatron, and Syzkaller). While these involve remote code sources, they are standard tools in the security research domain, and one source (Syzkaller) originates from a trusted organization.
- [DYNAMIC_EXECUTION]: The guides describe workflows that involve compiling source code at runtime (e.g., using
afl-clang-fast) to create instrumented binaries for testing. This is a common requirement for coverage-guided fuzzing. - [INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it is designed to ingest and analyze untrusted data generated during fuzzing sessions or network monitoring while maintaining powerful tool access.
- Ingestion points: Data enters the agent's context through fuzzer crash artifacts (
findings/default/crashes/), network traffic captures (.pcapfiles), and external API responses from tools likewfuzzorschemathesis. - Boundary markers: The skill does not specify the use of clear delimiters or instructions that warn the agent to ignore potentially malicious content embedded within crash reports or protocol data.
- Capability inventory: The agent has access to
Bash(command execution),WriteandEdit(file system modification), andWebFetch(network operations). - Sanitization: There is no explicit sanitization or validation logic defined for the external data before it is analyzed by the agent.
Audit Metadata