The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted source code across multiple programming languages including C, C++, Go, Rust, Java, Kotlin, Swift, PHP, Python, and Ruby. This creates an attack surface where malicious instructions embedded in the analyzed code or tool outputs could influence agent behavior.
Ingestion Points: User-supplied source files are analyzed using a Python-based analyzer tool (SKILL.md).
Capability Inventory: The skill involves calling numerous system compilers and disassemblers such as gcc, clang, rustc, javac, kotlinc, swiftc, ilspycmd, node, php, and ruby (SKILL.md, references/*.md).
Boundary Markers: The skill suggests structured output formats (e.g., --json), but lacks explicit delimiters or warnings to the agent regarding the potential for instructions embedded within the data it is analyzing.
Sanitization: No explicit sanitization of the analyzed content is described, although the instructions emphasize the use of isolated environments and stopping if tools are unavailable.
[COMMAND_EXECUTION]: The skill provides illustrative commands for the agent to execute for setup and analysis tasks.
Analysis Execution: Commands use uv run to execute a local Python analyzer script (SKILL.md).
Toolchain Setup: Instructions involve executing package management commands such as brew install, apt install, and dotnet tool install to prepare the analysis environment.
[PRIVILEGE_ESCALATION]: Setup guidelines for Linux platforms involve the use of sudo apt install to install necessary JDK components and toolchains (references/vm-compiled.md).
[PERSISTENCE]: The skill instructs the user or agent to modify shell configuration files (~/.zshrc or ~/.bashrc) to persist PATH environment variable updates for Homebrew-installed tools like OpenJDK and .NET (SKILL.md, references/vm-compiled.md).

constant-time-analysis