The Agent Skills Directory

[PROMPT_INJECTION] (HIGH): Vulnerable to Indirect Prompt Injection (Category 8). The skill is designed to ingest and process untrusted binary files using tools like strings, nm, and objdump. An attacker can craft a malicious binary with embedded instructions in symbol names, strings, or headers. Since the agent uses the output of these tools to guide its next steps and possesses powerful capabilities like Bash and Write, it could be tricked into performing harmful actions on the host system.
Ingestion points: Binary files (<binary>) provided as input to analysis tools.
Boundary markers: None present; tool output is processed directly by the agent.
Capability inventory: Bash tool (arbitrary command execution), Write tool (filesystem modification), and Python execution.
Sanitization: None; the skill lacks filtering or validation for tool output before it reaches the LLM context.
[COMMAND_EXECUTION] (HIGH): The skill provides the agent with the ability to execute arbitrary bash commands and Python scripts. While intended for legitimate binary analysis (e.g., checksec, pwntools), this level of access is dangerous if the agent's logic is compromised by malicious input.
[EXTERNAL_DOWNLOADS] (MEDIUM): The skill relies on numerous specialized third-party tools such as pwntools, cwe_checker, ROPgadget, one_gadget, qira, and Triton. If these are not pre-installed from trusted sources, they represent a significant supply-chain risk and a vector for remote code execution during the environment setup phase.

binary-analysis