CVE-2026-5970: FoundationAgents MetaGPT code injection in check_solution

Overview

CVE-2026-5970 is a code injection weakness classified as CWE-74 (Injection). The record places it in the check_solution function of the HumanEvalBenchmark/MBPPBenchmark component of MetaGPT, a multi-agent framework maintained by FoundationAgents. The description says that a manipulation of this function results in code injection and that the attack may be initiated remotely. Details of the input path are not given in the record.

Impact

The NVD lists a CVSS base score of 9.8 with network attack vector, low complexity, no privileges and no user interaction, with high impact on confidentiality, integrity and availability. Other scores in the record are lower (for example 7.3 and 5.5 from VulDB or its CNA). The SSVC data in the record marks the technical impact as partial and the exploitation level as proof of concept.

Affected products and versions

FoundationAgents MetaGPT: versions 0.8.0 and 0.8.1 are listed explicitly.
Deepwisdom Metagpt: listed as affected up to and including 0.8.1.
Versions before 0.8.0 are not named in the record; their status is unknown.

Exploitation status

The description states that the exploit is now public and may be used. The CISA KEV catalog does not list this CVE in the record, so no active exploitation is confirmed by the sources consulted. The EPSS probability in the record is 0.00713 (percentile 0.53132), a value that should be reviewed against the date recorded for it.

Remediation and mitigation

The record lists a patch reference to pull request 1988. Whether it has been merged or released is not stated.
Because the description says the project had not responded, organizations should check the project repository for a released version newer than 0.8.1 before relying on any fix.
Until a fixed release is verified, restrict who can supply input to the benchmark components and avoid running them on untrusted content or on hosts with sensitive data.
Isolate benchmark and evaluation workloads from production credentials and networks.

Detection guidance

Inventory all MetaGPT installations and record their exact versions, including any deployment that uses the HumanEvalBenchmark or MBPPBenchmark modules.
Review application and system logs for unexpected process creation, outbound connections or file changes originating from the MetaGPT process.
Monitor for evaluation jobs that receive unusual or externally supplied input.

Limitations of the available information

The record does not describe the exact input that triggers the flaw, the required configuration, or a proof of concept for defenders to test. It does not state whether the patch in pull request 1988 is merged, tested or included in a release. The relationship between the Deepwisdom and FoundationAgents product names is not explained. The EPSS date is later than the publication date and its meaning is not explained.

Frequently asked questions

Which MetaGPT versions are affected by CVE-2026-5970?

The record lists versions 0.8.0 and 0.8.1 explicitly and states that the issue affects versions up to and including 0.8.1. Versions before 0.8.0 are not named, so their status is unknown.

Is this CVE being actively exploited?

The record says a public exploit exists, but this CVE is not listed in CISA KEV and the sources consulted do not confirm active exploitation.

Is there a fix available?

The record references pull request 1988 as a patch, but it does not state whether it is merged or released, and the description says the project had not responded. Check the project repository for a newer release.

This analysis was drafted with AI assistance from verified source data and automatically checked against it.