Security context

What an agent needs to avoid regressing past fixes and find the next vuln in this repo.

sinaptik-ai/pandas-ai

              main @ bbbb771
            

Fixes

CVEs

CRITICAL

Peak severity

27.3%

Coverage

Highlights

Sandbox Bypass: 4 prior fixes. Scrutinize any change in this area.

pandasai/__init__.py: most-fixed (5 issues). Treat as high-risk during review.

10 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.

Sandbox Escape: Attackers can bypass execution sandboxes to run arbitrary commands on the host by exploiting weaknesses in AST analysis or importing un-whitelisted submodules. Securing the execution environment from escaping via nested or non-standard imports is a critical recurring challenge.

Code Injection: Prompt expansion variables and dynamic LLM responses are executed via Python's eval and exec. Untrusted inputs injected into evaluation scopes can trigger remote code execution if builtins and local contexts are not restricted.

SQL Injection: User-supplied or LLM-generated SQL queries are run directly against underlying databases. Without strict dialect-specific parsing and structural AST verification, malicious payloads can slip past regex-based filter logic to leak database contents.