The vulnerability resides in the check_solution function of the MetaGPT project’s HumanEvalBenchmark/MBPPBenchmark component. An attacker can supply a manipulated input that causes the function to evaluate arbitrary code, resulting in code injection. This weakness is identified as CWE‑74 and CWE‑94. The impact includes the potential execution of unintended code in the context of the MetaGPT application, which could lead to unauthorized data access, data modification, or further compromise of the host system.

MetaGPT versions up to 0.8.1 are affected. The vulnerability is present in the FoundationAgents MetaGPT repository under the HumanEvalBenchmark/MBPPBenchmark component. Users running any of these versions are at risk.

The CVSS score of 6.9 reflects a moderate severity: an attacker can remotely trigger the vulnerability without authentication. The exploit is publicly documented on GitHub and other advisory sites, but EPSS data is unavailable and it is not listed in the CISA KEV catalog. The attack vector is inferred to be remote, likely via an API call or request to the check_solution endpoint. Exploitation requires the ability to influence the input passed to the function; no additional preconditions are documented.