Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB
CVE-2026-105135 Overview
CVE-2026-105135 is a code injection vulnerability in InternLM MindSearch 0.1.0. The flaw affects the ExecutionAction.run function in mindsearch/agent/graph.py, part of the Planner Agent component. Attackers can manipulate the inputs argument to inject and execute arbitrary code remotely. The exploit has been publicly disclosed, and the vendor did not respond to disclosure attempts. MindSearch is an open-source multi-agent framework used for web search and reasoning tasks powered by large language models.
Critical Impact
Remote attackers can execute arbitrary code without authentication or user interaction by injecting payloads into the Planner Agent inputs, compromising confidentiality, integrity, and availability of the host system.
Affected Products
- InternLM MindSearch 0.1.0
- mindsearch/agent/graph.py — ExecutionAction.run function
- Planner Agent component
Discovery Timeline
- 2026-10-04 - CVE-2026-105135 published to NVD
- 2026-10-06 - Last updated in NVD database
Technical Details for CVE-2026-105135
Vulnerability Analysis
The vulnerability resides in the ExecutionAction.run method inside mindsearch/agent/graph.py. The Planner Agent receives attacker-controlled data through the inputs argument and passes it into a code execution path without adequate sanitization. This allows arbitrary Python code to be executed in the context of the MindSearch process.
The weakness maps to CWE-74, Improper Neutralization of Special Elements in Output Used by a Downstream Component (Injection). Because MindSearch typically runs with the privileges of the service account hosting the agent, successful exploitation grants attackers full control over the agent's execution environment.
Multi-agent LLM frameworks are frequently exposed to untrusted inputs through chat, API, or web search interfaces. This exposure makes injection flaws in planner components particularly impactful.
Root Cause
The root cause is unsafe handling of input data destined for a dynamic execution context. The ExecutionAction.run function evaluates planner-generated or user-influenced content as code without validating or sandboxing it. When untrusted input reaches a dynamic evaluator such as exec or eval, any embedded Python statement runs with the agent's privileges.
Attack Vector
The attack is initiated over the network and requires no authentication or user interaction. An attacker submits a crafted request to a MindSearch deployment that exposes the Planner Agent interface. The injected payload is then processed by ExecutionAction.run, which executes the attacker's code on the server.
Technical details and a proof of concept have been published in a GitHub PoC Repository and documented in the VulDB CVE-2026-105135 entry.
Stop Remote Code Execution Before It Runs
Singularity™ Endpoint detects and blocks exploit attempts at machine speed, including zero-days and novel attack techniques.
Stop Exploit AttemptsDetection Methods for CVE-2026-105135
Indicators of Compromise
- Unexpected child processes spawned by the MindSearch Python process, including sh, bash, curl, wget, or python subprocesses.
- Outbound network connections from the MindSearch host to unfamiliar IPs or domains shortly after agent requests.
- Anomalous entries in MindSearch request logs containing Python syntax such as __import__, os.system, or subprocess keywords in the inputs field.
- New or modified files in the MindSearch working directory not tied to normal operation.
Detection Strategies
- Enable application-level logging of all inputs passed to the Planner Agent and alert on payloads containing executable Python constructs.
- Monitor process ancestry on hosts running MindSearch and flag Python processes spawning shells or network utilities.
- Deploy a web application firewall (WAF) in front of the MindSearch API to inspect requests for injection patterns.
Monitoring Recommendations
- Baseline normal MindSearch behavior and alert on deviations in CPU, memory, and outbound traffic.
- Forward application and host telemetry to a centralized analytics platform for correlation across agent requests, process events, and network flows.
- Review access logs for high-frequency requests from single sources targeting agent endpoints.
How to Mitigate CVE-2026-105135
Immediate Actions Required
- Remove MindSearch 0.1.0 from internet-exposed systems until a vendor patch is available.
- Place the service behind an authenticated reverse proxy and restrict access to trusted networks only.
- Audit MindSearch logs for suspicious inputs content and signs of prior exploitation.
- Run MindSearch under a dedicated low-privilege account with no write access outside its working directory.
Patch Information
No vendor patch is available at the time of publication. According to the VulDB CVE-2026-105135 advisory, the vendor was contacted but did not respond. Operators must apply compensating controls until InternLM publishes a fix. Monitor the MindSearch project repository and VulDB entry for updates.
Workarounds
- Modify ExecutionAction.run in mindsearch/agent/graph.py to validate inputs against an allow-list and avoid passing untrusted data to dynamic execution primitives such as exec or eval.
- Execute the Planner Agent inside a hardened sandbox or container with no network egress and read-only filesystem mounts.
- Enforce strict input schema validation at the API layer to reject requests containing Python keywords or shell metacharacters.
- Apply egress firewall rules to block outbound connections from the MindSearch host to untrusted destinations.
# Example container hardening for MindSearch
docker run --rm \
--read-only \
--tmpfs /tmp:rw,noexec,nosuid,size=64m \
--cap-drop=ALL \
--security-opt no-new-privileges \
--network=internal-only \
--user 10001:10001 \
mindsearch:0.1.0
Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB

