InternLM MindSearch 0.1.0 unauthenticated code injection vulnerability

The Planner Agent in InternLM MindSearch 0.1.0 has an unauthenticated code injection vulnerability, numbered CVE-2026-105135.

InternLM's open-source AI search agent MindSearch has been disclosed with a code injection vulnerability CVE-2026-105135. VulDB, as a CNA, published the record on October 4, 2026. There is currently a public PoC and no known fix. The vulnerability is located in the Planner Agent's ExecutionAction.run (mindsearch/agent/graph.py), where the component passes Python code strings generated by the language model directly to exec for execution. An attacker who can influence the model output can make the server run arbitrary code. VulDB gives a CVSS v3.1 score of 10.0 and CVSS v4.0 of 9.3. NVD has not yet analyzed or scored it on its own. The GitHub mirror is the unverified advisory GHSA-4535-w5r2-pmqp, which does not list a fixed version.