The Agent Skills Directory

Dynamic Code Generation and Execution: The skill is designed to author and execute scripts (e.g., Python or C) to verify security vulnerabilities. While executing dynamically generated code is a high-capability operation, the skill incorporates strict guardrails to mitigate risks:
Sandbox Isolation: All executions are required to run inside containerized or isolated environments, with a specific prohibition against running commands directly on the host terminal.
Private Build Shadows: Compilation and test execution must occur in a temporary, private directory (mktemp -d), ensuring the original target source remains read-only and preventing unintended side effects.
Network Restrictions: The instructions specify that the environment should not be exposed to the external internet, particularly for memory-safety verification.
Shell Command Execution: The skill uses shell commands for file management (mkdir, cp, os.replace) and for launching debuggers or emulators (QEMU, Docker). These commands are used to manage the testing lifecycle and are protected by the sandbox constraints mentioned above.
Indirect Prompt Injection Surface: The skill ingests untrusted data from the target_root (the codebase under test) and from external findings JSON files. There is a potential consideration where malicious instructions embedded in source code comments or data files could attempt to influence the agent's logic during script generation. However, the skill's focus on empirical verification (checking for actual crashes or sanitizer hits) provides a functional check against purely instruction-based manipulation.

mantis-reproduce