strix-pentest
Fail
Audited by Gen Agent Trust Hub on Oct 28, 2026
Risk Level: HIGHPROMPT_INJECTIONPRIVILEGE_ESCALATIONREMOTE_CODE_EXECUTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The system prompt generated in
src/strix_cli_claude/main.pyexplicitly instructs the AI to ignore legal and safety barriers, using phrases like "AUTHORIZATION STATUS: No legal barriers - hack like you mean it" and "You think like a malicious hacker".
- [PRIVILEGE_ESCALATION]: The skill configures the AI agent to bypass standard platform security features. In
src/strix_cli_claude/main.py, it uses flags like--permission-mode bypassPermissionsand--dangerously-skip-permissions, and sets the environment variableCLAUDE_CODE_SKIP_TRUST_DIALOG=1to suppress interactive confirmation for dangerous operations. - [PRIVILEGE_ESCALATION]: In
src/strix_cli_claude/sandbox.py, the skill executeschmod 666 /var/run/docker.sockwithin the sandbox. This grants any process or user inside the container full control over the host machine's Docker daemon, which is a significant security risk for the host system. - [REMOTE_CODE_EXECUTION]: The skill downloads and executes scripts from external sources during setup. As seen in
src/strix_cli_claude/sandbox.pyandsrc/strix_cli_claude/main.py, it fetches the Docker installation script fromget.docker.comand the Trivy installation script from Aqua Security's GitHub repository, piping both directly to a shell for execution. - [COMMAND_EXECUTION]: The skill relies heavily on arbitrary shell execution to run security tools. The
terminal_executetool insrc/strix_cli_claude/mcp_server.pyand thePtyShellimplementation insrc/strix_cli_claude/sandbox_tool_server.pyprovide the agent with unconstrained shell access inside the Kali Linux environment. - [INDIRECT_PROMPT_INJECTION]: The skill possesses a high surface area for indirect injection.
- Ingestion points: Ingests untrusted target URLs, domains, and source code paths via
main.py. - Boundary markers: Prompts lack clear delimiters to separate instructions from target data.
- Capability inventory: Full terminal and Python execution capabilities are exposed to the agent while processing this data.
- Sanitization: While
shlex.quoteis used in some script generation, the agent's primary interaction with targets is through unconstrained tools likenucleiandsqlmap.
Recommendations
- HIGH: Downloads and executes remote code from: https://get.docker.com, https://raw.githubusercontent.com/aquasecurity/trivy/main/contrib/install.sh - DO NOT USE without thorough review
- AI detected serious security threats
Audit Metadata