agentic-pentest

Fail

Audited by Gen Agent Trust Hub on Sep 19, 2026

Risk Level: HIGHREMOTE_CODE_EXECUTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill provides instructions to download code and frameworks from several GitHub repositories that are not affiliated with trusted organizations or well-known services.\n
  • Evidence includes git clone commands for: github.com/GreyDGL/PentestGPT, github.com/ZacharyZcR/Hexstrike, github.com/jorhelpszg/Viper, github.com/ddzzh777/PentestAgent, github.com/yuvaly0/AI-Infra-Guard, and github.com/ArturSpirin/AutoPWN.\n- [REMOTE_CODE_EXECUTION]: The skill automates the execution of setup scripts and dependency installations from these untrusted sources, which can result in arbitrary code execution on the host system.\n
  • Evidence: pip3 install -r requirements.txt for PentestGPT, ./setup.sh for HexStrike AI, and docker compose up for Viper.\n- [COMMAND_EXECUTION]: The skill utilizes the high-privilege Bash tool to perform complex security operations and automation, which could be exploited if the agent's reasoning process is manipulated.\n
  • Evidence: The allowed-tools list includes Bash, and the methodology describes a loop where the agent proposes and executes shell commands for reconnaissance and exploitation.\n- [INDIRECT_PROMPT_INJECTION]: The skill's architecture is designed to ingest data from external targets (e.g., web pages and API responses) and process them through an LLM reasoning loop. This creates an attack surface where an adversary can embed malicious instructions in target data to hijack the agent's behavior.\n
  • Ingestion points: WebFetch tool results and tool outputs from ffuf/nuclei (SKILL.md, payloads.md).\n
  • Boundary markers: The skill suggests using delimiters in the prime, but these are not foolproof (guides/agentic-pentest-playbook.md).\n
  • Capability inventory: The agent has access to Bash, file read/write, and network tools (SKILL.md).\n
  • Sanitization: Documentation mentions the need for sanitization, but specific implementation for all ingested data is not provided within the skill's logic.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Sep 19, 2026, 04:36 PM