AIKIDO-2026-710104
verl is vulnerable to Code Injection
81
High Risk
This Affects:
verlTL;DR
The Qwen3 XML tool parser in the agent-loop rollout converts tool-call parameter values taken from model-generated output. When a parameter is typed as an array or another non-primitive type and cannot be decoded as JSON, the parser passes the raw value straight to Python eval(). Model output such as __import__("os").system(...) is therefore executed as Python during parsing, letting untrusted or externally influenced model output run arbitrary code on the training or rollout host. The fix replaces eval() with ast.literal_eval(), which only parses Python literals and degrades non-literal values to a string.
Who does this affect?
You are affected if you are using a version that falls within the vulnerable range and you run agent-loop rollouts that parse Qwen3 tool calls whose parameter values come from untrusted or externally influenced model output.
Background info
verl is vulnerable to Code Injection in versions 0.7.1 - 0.8.0.
How to fix this
Upgrade the verl library to the patch version.
Are You Affected?
Connect your repositories to instantly see whether vulnerable or malicious packages exist in your codebase.
Free. No credit card required.
I consent to receiving marketing communications based on Aikido’s Privacy Policy.
SOC 2Compliant
ISO 27001CompliantISO 42001Compliant
FedRAMPImplementing© All Intel data is openly available and commercially licensed.