Eval Injection in Google langfun

Executive Summary

CVE-2026-75062 is an eval injection flaw in Google's langfun library. The default lf.query protocol improperly neutralizes directives in dynamically evaluated code, allowing remote attackers to execute arbitrary Python code without authentication. The issue occurs because model outputs are parsed using pyglove.coding.evaluate with sandboxing disabled, enabling attacker-controlled input to bypass security filters and run malicious code in the host application.

Useful 0 Not Useful 0

Detection Guidance

Check if your langfun library version is prior to 0.1.2 by running pip show langfun. If installed, inspect code using lf.query for exec() calls without sandboxing. Monitor for unexpected network activity or file modifications during model interactions.

Useful 0 Not Useful 0

Impact Analysis

This vulnerability allows attackers to execute arbitrary Python code in your application's context. They could read sensitive data, modify files, make unauthorized network requests, or perform other malicious actions with the same privileges as your application. The attack is silent since malicious responses still return valid schema objects, making detection difficult.

Useful 0 Not Useful 0

Compliance Impact

This vulnerability likely violates compliance requirements for data protection and security. It enables unauthorized code execution which could lead to data breaches, unauthorized access to sensitive information, or system modifications. Organizations using affected langfun versions may fail GDPR's security principle (Article 32) and HIPAA's security rule (45 CFR Part 164) for protecting electronic protected health information.

Useful 0 Not Useful 0

Mitigation Strategies

Upgrade langfun to version 0.1.2 or later. If upgrading is not possible, disable the lf.query protocol or implement a strict input validation layer to sanitize model outputs before execution. Avoid using untrusted inputs in model prompts.

Useful 0 Not Useful 0