llm-as-computer
Fail
Audited by Gen Agent Trust Hub on Oct 13, 2026
Risk Level: HIGHDATA_EXFILTRATIONCOMMAND_EXECUTIONDYNAMIC_EXECUTIONEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTION
Full Analysis
- [DATA_EXFILTRATION]: The update procedure in
SKILL.mdincludes shell commands to read a sensitive credential file located at/mnt/project/GitHub.env. The extractedGH_TOKENis then transmitted in the headers of an outgoingcurlrequest to the GitHub API. - [COMMAND_EXECUTION]: The skill performs system-wide modifications by installing the Mojo toolchain and related Python packages using
uv pip installwith the--systemand--break-system-packagesflags, which can destabilize the host's package management and environment isolation. - [DYNAMIC_EXECUTION]: The skill performs runtime compilation of its own distributed source code using the
mojo buildcommand to generate an executable binary (percepta_exec), which is then invoked via subprocess calls. - [EXTERNAL_DOWNLOADS]: The skill fetches updated scripts and source files from the author's public GitHub repository (
oaustegard/llm-as-computer). - [REMOTE_CODE_EXECUTION]: The update logic involves downloading executable scripts and source files from a remote repository and subsequently executing or compiling them, allowing for remote modification of the skill's logic.
Recommendations
- AI detected serious security threats
Audit Metadata