# VEXAIoT: Autonomous Multi-Agent LLM Framework Automates End-to-End IoT Vulnerability Discovery and Exploitation (95% Success Rate)

> Academic researchers (Swinea, Aryal, Praharaj, Gupta) published VEXAIoT, a two-agent LLM framework (ChatGPT 5.1 thinking) that chains Nmap reconnaissance, Searchsploit/Exploit-DB lookups, Bettercap, and John the Ripper into autonomous, dependency-aware IoT attack plans. Tested against IoTGoat and Metasploitable2 across 260 executions covering 10 OWASP IoT Top 10 attack scenarios (hardcoded password cracking, MiniUPnP backdoor abuse, XSS injection, developer backdoor access, malicious firmware update execution, DNS DoS, database PII extraction, MitM credential interception, log erasure, and RCE), it achieved a 95% overall success rate (94.5% on IoTGoat, 96.7% on Metasploitable2) with most attacks completing in under two minutes and low token overhead (460-1,480 tokens/attack).

- **Published:** 2026-07-13T00:00:00Z
- **Last reviewed:** 2026-07-13T00:00:00Z
- **Canonical:** https://intel.threadlinqs.com/threat/TL-2026-1261
- **Card image:** https://intel.threadlinqs.com/og/threat/threat-intel-vexaiot-llm-agent-framework-tl-2026-1261.webp?v=62707487 (Threadlinqs threat intel: VEXAIoT Multi-Agent LLM Attacks on IoT: Analysis & Hunting (TL-2026-1261) — Medium severity, 24 IOCs, 9 detection rules.)
- **ID:** TL-2026-1261
- **Severity:** MEDIUM
- **Category:** TOOL
- **Status:** ACTIVE
- **Detections:** 9 (1 shown free below) · **IOCs:** 24
- **CVEs:** CVE-2011-2523, CVE-2004-2687, CVE-2019-12108, CVE-2019-12109, CVE-2019-12110, CVE-2019-12111, CVE-2020-28951

## Description

VEXAIoT ("Vulnerability EXploitation using AI Agents") is a research framework published on arXiv (2607.09653, submitted 2026-07-10, CC BY 4.0) by Katherine Swinea, Kshitiz Aryal, Lopamudra Praharaj, and Maanak Gupta, first reported publicly by Cyber Security News on 2026-07-13. The system decomposes offensive IoT security assessment into a modular multi-agent pipeline: reconnaissance -> vulnerability mapping -> attack planning -> execution -> validation.

Architecturally, VEXAIoT uses two cooperating agents built on ChatGPT 5.1 ("thinking" reasoning mode): a Vulnerability Detection Agent that runs Nmap-driven reconnaissance to enumerate hosts, open ports, exposed services, and software versions, then correlates discovered software against known CVEs via Searchsploit and Exploit-DB; and an Attack Execution Agent that consumes the detection agent's findings, selects an appropriate offensive tool (Nmap NSE scripts, Searchsploit-sourced PoC exploits, Bettercap for MITM/traffic interception, John the Ripper for password cracking), generates the exact command sequence, executes it, and validates success using LLM-based interpretation of tool output. When a command fails, the LLM revises its approach using error-message feedback and retry logic rather than following a rigid, pre-scripted playbook -- the paper's central claim of novelty versus prior CTF-focused LLM pentesting tools such as PentestGPT, AutoAttacker, and AutoPentester.

The authors deliberately used functional, non-malicious tool-description framing in their prompts specifically to reduce LLM safety-refusal behavior during offensive tasking -- a prompt-engineering technique with dual-use implications for real-world adversarial automation, since it demonstrates a working method to lower a commercial LLM's resistance to producing exploit commands.

Evaluation targeted two intentionally vulnerable testbeds: IoTGoat (OWASP's deliberately insecure OpenWrt-based IoT firmware, containing MiniUPnP 2.1 DoS flaws CVE-2019-12108/12109/12110/12111 and libuci use-after-free CVE-2020-28951) and Metasploitable2 (a classic vulnerable Linux distribution whose exploited services in this study align with the well-known vsftpd 2.3.4 backdoor CVE-2011-2523 and distcc remote command execution CVE-2004-2687). Across 260 total execution runs mapped to 10 OWASP IoT Top 10 categories, VEXAIoT achieved 247/260 (95.0%) successful exploitations: 189/200 (94.5%) on IoTGoat and 58/60 (96.7%) on Metasploitable2. Seven of ten scenarios (XSS injection, developer backdoor access, malicious update execution, database PII extraction, log erasure, and RCE) reached 100% success; MiniUPnP backdoor abuse and DNS DoS scenarios lagged at 80% due to command-syntax generation errors and model refusals. Parallelized execution completed all 10 attack scenarios in ~3 minutes 50 seconds versus ~8 minutes 31 seconds sequentially. Of the 13 failures, the paper attributes 5 to model refusals, 5 to LLM-generated command syntax errors, and 1 to model hallucination (2 unaccounted/other in the reported breakdown).

No CVE was assigned to VEXAIoT itself (it is a research tool/methodology, not a vulnerability), and no threat-actor adoption or in-the-wild use has been reported. The risk profile is that of a capability-lowering dual-use tool: it demonstrates that commodity multi-agent LLM tooling can autonomously chain reconnaissance-to-exploitation against IoT/embedded Linux targets with high reliability and low cost, shortening the skill and time barrier previously required for manual IoT penetration testing or malicious exploitation. The authors explicitly scope the work to isolated, authorized testbeds and recommend production/real-world deployment safeguards: strict access controls, comprehensive action logging, mandatory human approval gates for sensitive actions, and technical guardrails preventing agents from reaching hosts outside an approved test scope. They acknowledge that autonomous exploit execution at scale requires stronger command validation and safety controls before any responsible real-world use, and note the study does not reflect production-network complexity (patched software, IDS/IPS, network segmentation, device heterogeneity) that would degrade success rates outside a lab.

## MITRE ATT&CK

- T1595.002 Vulnerability Scanning
- T1592.002 Software
- T1588.005 Exploits
- T1588.002 Tool
- T1190 Exploit Public-Facing Application
- T1078.001 Default Accounts
- T1059.004 Unix Shell
- T1505.003 Web Shell
- T1068 Exploitation for Privilege Escalation
- T1685.006 Clear Linux or Mac System Logs
- T1110.002 Password Cracking
- T1557 Adversary-in-the-Middle
- T1552.001 Credentials In Files
- T1046 Network Service Discovery
- T1082 System Information Discovery
- T1005 Data from Local System
- T1557.003 DHCP Spoofing
- T1071 Application Layer Protocol
- T1041 Exfiltration Over C2 Channel
- T1498 Network Denial of Service
- T1495 Firmware Corruption
- T1565 Data Manipulation

## Free detection (Splunk SPL, Microsoft KQL, Sigma)

**vsftpd 2.3.4 Trojaned Backdoor Login Trigger and distccd Unauthenticated RCE** (`TL-2026-1261-SPL-002`)

Detects the two Metasploitable2-class exploitation primitives VEXAIoT automated during post-exploitation: an FTP login attempt containing the vsftpd 2.3.4 smiley-face backdoor trigger sequence (CVE-2011-2523) and distccd spawning a shell interpreter as a child process consistent with unauthenticated command execution (CVE-2004-2687).

- **Languages:** Splunk SPL, Microsoft KQL, Sigma
- **ATT&CK:** T1190, T1078.001, T1059.004
- **Data sources:** linux_secure, ftp_logs, sysmon, DeviceProcessEvents, DeviceNetworkEvents
- **Severity:** critical
- **Confidence:** high

### Splunk SPL

```spl
index=network sourcetype IN (linux_secure, ftp_logs, sysmon) (CommandLine="*/bin/sh*" ParentImage="*distccd*") OR (Message="*USER*:)*" OR Message="*USER *)*" service="ftp")
| eval detection_reason=case(match(ParentImage, "distccd"), "distccd_rce_CVE-2004-2687", match(Message, "\)\)\)|:\)"), "vsftpd_backdoor_trigger_CVE-2011-2523", 1=1, "other")
| where detection_reason!="other"
| table _time host user CommandLine ParentImage Message detection_reason
```

### Microsoft KQL

```kql
let distcc_rce = DeviceProcessEvents
| where Timestamp > ago(24h)
| where InitiatingProcessFileName has "distccd" and ProcessCommandLine has_any ("/bin/sh", "id;")
| project Timestamp, DeviceName, ProcessCommandLine, InitiatingProcessFileName, DetectionReason="distccd_rce_CVE-2004-2687";
let vsftpd_trigger = DeviceNetworkEvents
| where Timestamp > ago(24h)
| where RemotePort == 21 and AdditionalFields has ":)"
| project Timestamp, DeviceName, ProcessCommandLine=tostring(AdditionalFields), InitiatingProcessFileName="vsftpd", DetectionReason="vsftpd_backdoor_trigger_CVE-2011-2523";
distcc_rce
| union vsftpd_trigger
```

### Sigma

```yaml
title: distccd Unauthenticated Command Execution (CVE-2004-2687)
id: 7c3a1e2a-1f4b-4b7e-9a0e-1d2c3b4a5f02
status: experimental
description: Detects distccd spawning a shell interpreter as a direct child process, consistent with unauthenticated remote command execution via CVE-2004-2687, as automated by VEXAIoT's Attack Execution Agent against Metasploitable2-class targets.
references:
    - https://nmap.org/nsedoc/scripts/distcc-cve2004-2687.html
    - https://arxiv.org/abs/2607.09653
author: Threadlinqs Team
date: 2026-07-13
tags:
    - attack.initial-access
    - attack.t1190
    - attack.execution
    - attack.t1059.004
    - cve.2004.2687
logsource:
    category: process_creation
    product: linux
detection:
    selection:
        ParentImage|endswith: '/distccd'
        CommandLine|contains:
            - '/bin/sh'
            - 'id;'
    condition: selection
falsepositives:
    - Legitimate distributed compilation jobs (distcc used as intended within a trusted, access-controlled build network)
level: high
```

**Known false positives:**

- Legitimate distcc build-farm traffic from trusted internal build servers
- FTP clients that happen to include a colon-parenthesis sequence in a legitimate username by coincidence

**Tuning notes:** Restrict distccd child-shell alerting to hosts outside the approved build-network CIDR; require distccd hardening (--allow flag) or full decommission where not operationally required.

### 8 more detection rules (Splunk SPL, Microsoft KQL, Sigma), unlock with TL Creds

- OpenWrt [indicator] / MiniUPnP Crash-Loop Exploitation (IoTGoat-class Devices) (locked: unlock with TL Creds)
- Hardcoded/Default Credential Access Followed by Local Password-Cracking Utility Execution (locked: unlock with TL Creds)
- Rapid Automated Log-Clearing Immediately Following Exploit/RCE Activity (locked: unlock with TL Creds)
- MiniUPnP / [indicator] / [indicator] Rapid Multi-Service Scan Preceding Exploit (Exploitation Phase) (locked: unlock with TL Creds)
- MiniUPnP Backdoor/Service Abuse and Developer-Backdoor Interface Access (Post-Exploitation Phase) (locked: unlock with TL Creds)
- Adversary-in-the-Middle ARP/DHCP Spoofing and Credential Interception (C2/Lateral Movement Phase) (locked: unlock with TL Creds)
- Agentic Recon-to-Exploit Command Sequencing (VEXAIoT-class Chaining) (locked: unlock with TL Creds)
- [indicator] 2.3.4 Backdoor Shell Established on TCP/6200 (locked: unlock with TL Creds)

Sign up free to unlock the detection rules: https://threadlinqs.com/?modal=signup

## Sources

- [VEXAIoT Automates IoT Reconnaissance](https://cybersecuritynews.com/vexaiot-automates-iot-reconnaissance/)
- [VEXAIoT: Autonomous IoT Vulnerability EXploitation using AI Agents (arXiv:2607.09653)](https://arxiv.org/abs/2607.09653)
- [VEXAIoT arXiv HTML full text](https://arxiv.org/html/2607.09653v1)
- [AI agents automate IoT attacks with 95% success](https://thearabianpost.com/ai-agents-automate-iot-attacks-with-95-success/)
- [VSFTPD 2.3.4 Backdoor Command Execution (Rapid7 module)](https://www.rapid7.com/db/modules/exploit/unix/ftp/vsftpd_234_backdoor/)
- [distcc-cve2004-2687 NSE script — Nmap Scripting Engine documentation](https://nmap.org/nsedoc/scripts/distcc-cve2004-2687.html)
- [OWASP/IoTGoat — deliberately insecure OpenWrt-based firmware](https://github.com/OWASP/IoTGoat)
- [IoTGoat challenges wiki](https://github.com/scriptingxss/IoTGoat/wiki/IoTGoat-challenges)
- [PentestGPT: An LLM-empowered Automatic Penetration Testing Tool (arXiv:2308.06782)](https://arxiv.org/abs/2308.06782)

## Full data

The other 8 detection rules (Splunk SPL / Microsoft KQL / Sigma) unlock with TL Creds; a free account includes daily creds: https://threadlinqs.com/?modal=signup
Indicator values unlock with TL Creds, at a lower price on Red and Purple; AI agents can query rules and IOCs through the Threadlinqs MCP server (Purple): https://intel.threadlinqs.com/mcp

Canonical: https://intel.threadlinqs.com/threat/TL-2026-1261
