Security context

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

agenta-ai/agenta

              main @ b6fafc3
            

134

Fixes

CVEs

CRITICAL

Peak severity

32.3%

Coverage

Highlights

Auth Bypass: 82 prior fixes. Scrutinize any change in this area.

agenta-cli/agenta/sdk/middleware/auth.py: most-fixed (12 issues). Treat as high-risk during review.

78 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.

Server-Side Template Injection: Unsafe template rendering in the API server evaluator allowed remote attackers to execute arbitrary code. The fix replaced standard Jinja2 rendering with a SandboxedEnvironment. This risk is tracked under CVE-2026-27961.

Sandbox Escape: Improper sandboxing of user-supplied code via whitelisted libraries in the custom code evaluation runtime allowed authenticated users to bypass RestrictedPython boundaries and execute arbitrary commands. The feature was ultimately removed to prevent sandbox escapes. This risk is tracked under CVE-2026-27952.

Auth Bypass: The SDK middleware suffered from multiple authorization bypass issues due to caching misconfigurations, incorrect handling of falsy policy objects, unauthenticated route leakage, and incorrect extraction of project identifiers from requests.