Security context
What an agent needs to avoid regressing past fixes and find the next vuln in this repo.
business-science/ai-data-science-team
master @ 4ffeb7f
4
Fixes
0
CVEs
HIGH
Peak severity
Highlights
Arbitrary Code Execution: 3 prior fixes. Scrutinize any change in this area.
ai_data_science_team/agents: most-fixed (1 issue). Treat as high-risk during review.
4 high-severity fixes in this history; regressions here are high-impact.
Recurring patterns
The bug types that recur here, drawn from past fixes, not open vulnerabilities.
Remote Code Execution: LLM-generated agent code was executed directly within the host process. This allowed untrusted code to run with the full privileges of the application, requiring a migration to a sandboxed subprocess.
Arbitrary Code Execution: Unsafe in-process evaluation of untrusted data-wrangling code generated by language models posed an immediate system compromise risk before it was isolated into a restricted subprocess.
Arbitrary Code Execution: The data cleaning agent evaluated arbitrary LLM-generated code in the main process, necessitating a sandboxed subprocess execution environment with restricted builtins and blocked imports.