
🚨 #ALERT — DEEPTUTOR AI AGENT TOOL PATH CAN REACH HOST SHELL COMMAND EXECUTION September 30, 2026 DISCLOSED BY: Security researcher / MITRE CVE record PRODUCT: DeepTutor — TutorBot / AI agent tooling CVE: CVE-2026-51870 — CVSS 9.8 AFFECTED VERSIONS: CVE record identifies DeepTutor v1.4.0. Project release notes state v1.4.1 “locked down” the TutorBot tool sandbox; operators should move to a current release rather than remain on v1.4.0. IMPACT: A user-controlled TutorBot interaction can cause the LLM-backed tool layer to pass generated commands into ExecTool.execute, reaching the host shell. The published validation path is: User message → TutorBot WebSocket → LLM → ExecTool → system shell. EXPLOITATION STATUS: NO TRUSTED CONFIRMATION OF IN-THE-WILD EXPLOITATION IDENTIFIED. PUBLIC PoC / REPRODUCTION: AVAILABLE. IMPORTANT: The published reproducer assumes a reachable TutorBot with Shell/Exec capability enabled. This does not establish that every DeepTutor deployment is remotely exploitable under identical conditions. Forensic triage: Review TutorBot/WebSocket sessions, tool-call traces, ExecTool activity, and unexpected child processes or shell commands originating from the DeepTutor service. URGENT ACTION: Do not expose v1.4.0 deployments. Upgrade to a current release, disable Shell/Exec tools unless required, sandbox agent execution, and run tool processes with minimal privileges. SEAL LINK: DIRECT — this demonstrates an agent execution-boundary failure where user-controlled language can drive a tool path into host shell execution. It should not be mislabeled as a sandbox escape; the exposed execution path itself is the issue. SOURCE: https://github.com/HKUDS/DeepTutor/issues/506 VALIDATION: https://github.com/HKUDS/DeepTutor/issues/518 CVE: https://nvd.nist.gov/vuln/detail/cve-2026-51870?utm_source=chatgpt.com CONFIDENCE: VERY HIGH for the command-execution path and public reproduction. INSUFFICIENT for any claim of confirmed real-world exploitation. #CyberSecurity #ThreatIntel #NØØT #DeepTutor #AIAgents #AgenticAI #AISecurity #RCE #CommandInjection #LLMSecurity #AgentSecurity #Sandboxing #CVE_2026_51870
Post summary
The text is an alert disclosing a critical command execution vulnerability (CVE-2026-51870) in DeepTutor v1.4.0, highlighting the availability of a public PoC and providing specific remediation steps, including upgrading to v1.4.1 and applying security workarounds.

