strix
Fail
Audited by Gen Agent Trust Hub on Nov 18, 2026
Risk Level: HIGHREMOTE_CODE_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [REMOTE_CODE_EXECUTION]: The
scripts/install.shscript downloads an installer fromhttps://strix.ai/installand executes it directly usingbash. While the script contains a conditional checksum verification step, the required hash variable (STRIX_INSTALL_SHA256) is not populated within the skill's code, meaning the integrity of the downloaded script is not validated by default. - [EXTERNAL_DOWNLOADS]: The skill performs several network requests to fetch external components:
scripts/install.shdownloads a shell script fromhttps://strix.ai/install.- The installation process pulls a Docker image (
ghcr.io/usestrix/strix-sandbox:0.1.13) from the GitHub Container Registry. SKILL.mdandreferences/commands.mdencourage the download of installation scripts directly from thestrix.aidomain.- [INDIRECT_PROMPT_INJECTION]: The skill's primary purpose is to process untrusted external data (GitHub repositories, staging URLs, and local codebases) through an LLM for security analysis, which creates a vulnerability to instructions embedded in that data.
- Ingestion points: The
--target(or-t) argument inscripts/run-scan.shandscripts/ci-scan.shaccepts arbitrary user-controlled paths and URLs. - Boundary markers: The skill's scripts and instructions do not include boundary markers or system-level directives to prevent the agent from obeying instructions hidden within the target files or web pages.
- Capability inventory: The skill has access to
Bash,Read,Write, andWebFetchtools. Thestrixtool itself has capabilities for browser automation and shell execution within a sandbox. - Sanitization: No sanitization or filtering of content retrieved from the targets is performed before it is analyzed by the LLM.
- [COMMAND_EXECUTION]: The skill provides multiple wrapper scripts (
scripts/run-scan.sh,scripts/ci-scan.sh, andscripts/install.sh) that execute shell commands, includingdocker,strix, andcurl. While intended for the tool's operation, these scripts handle user-provided arguments like targets and scan modes.
Recommendations
- AI detected serious security threats
Audit Metadata