# APT31 Weaponizes Google Gemini AI for Automated Cyberattack Planning

> Google Threat Intelligence Group (GTIG) published its Q4 2025 AI Threat Tracker revealing that APT31 (Violet Typhoon/Zirconium/Judgment Panda), a PRC-backed cyber espionage group sanctioned by the US in March 2024, weaponized Google Gemini AI with Hexstrike MCP red-teaming tooling to automate vulnerability analysis, generate targeted penetration testing plans, and conduct reconnaissance against specific US-based targets. The report also documents AI-augmented operations by APT42 (Iran), UNC2970 (DPRK), APT41 (PRC), UNC795 (PRC), and UNC6418, alongside new threats including HONESTCUE AI-integrated malware, COINBAIT AI-generated phishing kits, model extraction/distillation attacks against Gemini, and the Xanthorox underground AI-for-cybercrime service built on jailbroken commercial APIs via MCP servers.

- **Published:** 2026-02-15T04:55:00Z
- **Last reviewed:** 2026-02-15T04:55:00Z
- **Canonical:** https://intel.threadlinqs.com/threat/TL-2026-0085
- **Card image:** https://intel.threadlinqs.com/og/threat/threat-intel-apt31-gemini-ai-misuse-tl-2026-0085.webp?v=b192bbc3 (Threadlinqs threat intel: APT31 Abuses Google Gemini and Hexstrike: Detection & IOCs (TL-2026-0085) — Critical severity, 43 IOCs, 9 detection rules.)
- **ID:** TL-2026-0085
- **Severity:** CRITICAL
- **Category:** APT
- **Status:** SUPERSEDED
- **Actor:** APT31 (China)
- **Detections:** 9 (1 shown free below) · **IOCs:** 43 (3 shown free below)
- **CVEs:** CVE-2025-7775, CVE-2025-53770, CVE-2025-53771

## Description

APT31 Weaponizes Google Gemini AI for Automated Vulnerability Analysis & Cyberattack Planning

Executive Context:

On February 12, 2026, Google Threat Intelligence Group (GTIG) published its latest AI Threat Tracker report titled 'Distillation, Experimentation, and (Continued) Integration of AI for Adversarial Use,' documenting the escalating weaponization of AI tools by state-sponsored threat actors during Q4 2025. The report represents the most comprehensive public disclosure to date of how nation-state APT groups are operationalizing generative AI across the full attack lifecycle.

The headline finding centers on APT31, a PRC Ministry of State Security (MSS)-affiliated cyber espionage group, which employed a 'highly structured approach' to weaponize Google's Gemini AI chatbot. APT31 prompted Gemini with fabricated expert cybersecurity personas and integrated Hexstrike — an open-source red-teaming tool built on the Model Context Protocol (MCP) — to automate the analysis of remote code execution (RCE) vulnerabilities, web application firewall (WAF) bypass techniques, and SQL injection attack vectors against specific US-based targets.

APT31 Background:

APT31, also tracked as Violet Typhoon (Microsoft), Zirconium (Microsoft legacy), Judgment Panda (CrowdStrike), Bronze Vinewood (Secureworks), and Red Keres (PwC), is attributed to China's Ministry of State Security (MSS), specifically the Hubei State Security Department based in Wuhan. The group has been active since at least 2010 and targets government entities, international financial organizations, aerospace and defense companies, technology firms, and telecommunications providers worldwide.

In March 2024, the US Department of Justice indicted seven APT31 members (Ni Gaobin, Weng Ming, Cheng Feng, Peng Yaowen, Sun Xiaohui, Xiong Wang, and Zhao Guangzong) for a 14-year campaign targeting US critical infrastructure, political campaigns, and international organizations. The US Treasury Department's OFAC simultaneously sanctioned the Wuhan Xiaoruizhi Science and Technology Company (Wuhan XRZ) as a front company for APT31 operations.

Hexstrike MCP Integration:

Hexstrike is an open-source AI-powered penetration testing framework built on the Model Context Protocol (MCP) that enables AI models including Gemini to orchestrate 150+ security tools for network scanning, vulnerability scanning, reconnaissance, and penetration testing. Originally designed for ethical hackers and bug bounty hunters, Hexstrike was released in mid-August 2025 and quickly adopted by malicious actors. APT31's use of Hexstrike with Gemini represents one of the first documented cases of a nation-state APT group integrating agentic AI tooling into offensive cyber operations.

The integration automated intelligence gathering to identify technological vulnerabilities and organizational defense weaknesses against specific US-based targets. Google's GTIG chief analyst John Hultquist noted: 'We anticipate that China-based actors in particular will continue to build agentic approaches for cyber offensive scale.'

Additional Threat Actors Documented:

1. APT42 (Iran/IRGC) — Used Gemini for reconnaissance, social engineering augmentation, hyper-personalized phishing lure generation, rapport-building phishing, and offensive tooling development including malware debugging and code generation.

2. UNC2970 (DPRK/RGB) — Used Gemini to synthesize OSINT and profile high-value targets in the defense and cybersecurity sectors, mapping technical job roles and salary information to create tailored phishing personas for Operation Dream Job-style campaigns.

3. APT41 (PRC/Winnti) — Leveraged Gemini for knowledge synthesis, real-time troubleshooting, and code translation to accelerate malicious tooling development.

4. UNC795 (PRC) — Engaged Gemini multiple days per week across entire attack lifecycle; attempted to create AI-integrated code auditing capability demonstrating interest in agentic AI utilities.

5. UNC6418 (Unattributed) — Used Gemini for targeted intelligence gathering (credentials, emails) that directly preceded phishing campaigns targeting Ukraine and the defense sector.

6. Temp.HEX (PRC) — Used Gemini and other AI tools to compile detailed intelligence on specific individuals in Pakistan and separatist organizations in multiple countries.

New Malware & Tooling:

1. HONESTCUE — Novel AI-integrated downloader/launcher that sends prompts to Gemini's API and receives C# source code, which is compiled and executed in memory via .NET CSharpCodeProvider (fileless execution). Uses Discord CDN for payload hosting. Represents proof-of-concept of AI-outsourced malware functionality.

2. COINBAIT — AI-generated phishing kit built using Lovable AI platform, masquerading as cryptocurrency exchange (Coinbase) for credential harvesting. Linked to UNC5356 financially motivated cluster. Uses React SPA architecture with Supabase backend.

3. Xanthorox — Underground 'AI-for-cybercrime' toolkit advertised as custom AI but actually built on jailbroken commercial APIs (including Gemini) via MCP servers using Crush, Hexstrike AI, LibreChat-AI, and Open WebUI.

Model Extraction Attacks:

GTIG and Google DeepMind identified increased 'distillation attacks' — model extraction attempts targeting Gemini's reasoning capabilities. One campaign used 100,000+ prompts to attempt reasoning trace coercion. The attacks represent IP theft targeting proprietary model logic for cheaper replication.

AI-Enabled ClickFix Campaigns:

Threat actors abused public sharing features of AI services (Gemini, ChatGPT, CoPilot, DeepSeek, Grok) to host malicious ClickFix social engineering instructions, distributing ATOMIC macOS information stealer. First observed use of AI service public sharing as trusted domain hosting.

Strategic Implications:

1. The patch gap is widening — AI-powered vulnerability discovery gives attackers speed advantage over defenders who take weeks to deploy patches.
2. Agentic AI is the 'next shoe to drop' — autonomous multi-step offensive operations with minimal human oversight are becoming reality.
3. Traditional detection tells (poor grammar, cultural mismatches) are being erased by LLM-generated content.
4. MCP-based offensive tooling creates a new attack surface category — legitimate security tools repurposed for nation-state operations.
5. Underground AI services will increasingly rely on jailbroken commercial APIs rather than custom models.

## MITRE ATT&CK

- T1591 Gather Victim Org Information
- T1593 Search Open Websites/Domains
- T1589 Gather Victim Identity Information
- T1595 Active Scanning
- T1587 Develop Capabilities
- T1588 Obtain Capabilities
- T1608 Stage Capabilities
- T1583 Acquire Infrastructure
- T1566 Phishing
- T1190 Exploit Public-Facing Application
- T1189 Drive-by Compromise
- T1059 Command and Scripting Interpreter
- T1204 User Execution
- T1203 Exploitation for Client Execution
- T1027 Obfuscated Files or Information
- T1620 Reflective Code Loading
- T1036 Masquerading
- T1055 Process Injection
- T1056 Input Capture
- T1555 Credentials from Password Stores
- T1082 System Information Discovery
- T1005 Data from Local System
- T1119 Automated Collection
- T1567 Exfiltration Over Web Service
- T1496 Resource Hijacking
- T1585 Establish Accounts
- T1199 Trusted Relationship
- T1548 Abuse Elevation Control Mechanism
- T1539 Steal Web Session Cookie
- T1587.004 Exploits
- T1595.002 Vulnerability Scanning
- T1059.006 Python
- T1027.002 Software Packing
- T1090.003 Multi-hop Proxy
- T1547.001 Registry Run Keys / Startup Folder
- T1041 Exfiltration Over C2 Channel
- T1110.003 Password Spraying
- T1685 Disable or Modify Tools
- T1021.002 SMB/Windows Admin Shares
- T1584.005 Botnet

## Free detection (Splunk SPL)

**APT31 Gemini AI API Abuse from Non-Browser Process** (`TL-2026-0085-SPL-001`)

Detects connections to Google Gemini API (generativelanguage.googleapis.com) from non-browser processes such as PowerShell, Python, cmd, wscript, or cscript. APT31 used scripted Gemini API calls with expert cybersecurity persona prompts to automate vulnerability analysis, WAF bypass development, and targeted penetration testing plans against US organizations. HONESTCUE malware also queries Gemini API from non-browser contexts for fileless code generation.

- **Language:** Splunk SPL
- **ATT&CK:** T1587.004, T1059.006, T1059.003
- **Data sources:** proxy, dns_query, sysmon
- **Severity:** high
- **Confidence:** high

### Splunk SPL

```spl
index=proxy OR index=dns OR index=sysmon (dest="generativelanguage.googleapis.com" OR query="generativelanguage.googleapis.com" OR QueryName="generativelanguage.googleapis.com") NOT (process_name IN ("chrome.exe","msedge.exe","firefox.exe","brave.exe","opera.exe","safari","chromium") OR Image IN ("*\\chrome.exe","*\\msedge.exe","*\\firefox.exe","*\\brave.exe")) | eval risk_process=case(match(process_name,"(?i)(powershell|pwsh|python|python3|cmd|wscript|cscript|curl|wget|node)"),"HIGH",match(process_name,"(?i)(java|dotnet|go)"),"MEDIUM",1=1,"LOW") | stats count as api_calls earliest(_time) as first_seen latest(_time) as last_seen values(process_name) as processes values(src_ip) as source_ips by dest,risk_process | where api_calls>=1 | eval alert_msg="APT31-style Gemini API abuse: ".mvjoin(processes,",")." made ".api_calls." calls to ".dest | sort -risk_process,-api_calls
```

**Known false positives:**

- Legitimate AI development tools and IDEs using Gemini API
- CI/CD pipelines with AI integration
- Authorized security testing tools

**Tuning notes:** Whitelist known development workstations and CI/CD service accounts. Focus on unexpected processes (wscript, cscript, cmd) making Gemini API calls. Extend dest list to include api.openai.com, api.anthropic.com, api-inference.huggingface.co for broader AI API abuse coverage.

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

- PromptFlux Self-Rewriting AI Malware Behavioral Pattern (locked: unlock with TL Creds)
- APT31 ORB Proxy Multi-Hop C2 via Compromised Residential IPs (locked: unlock with TL Creds)
- AI Model Extraction via High-Volume Structured API Queries (locked: unlock with TL Creds)
- SharePoint CVE-2025-53770 Deserialization RCE Exploitation (locked: unlock with TL Creds)
- AI Safety Guardrail Bypass via Expert Persona Fabrication (locked: unlock with TL Creds)
- MCP Server Process Execution - HexStrike and Offensive AI Tooling (locked: unlock with TL Creds)
- HexStrike MCP Rapid Security Tool Chaining (locked: unlock with TL Creds)
- HONESTCUE CSharpCodeProvider In-Memory Compilation from Network Code (locked: unlock with TL Creds)

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

## Sample indicators (3 of 43)

Defanged, TLP:CLEAR.

- Domain: `lovable[.]app`
- URL: `hxxps://github[.]com/0x4m4/hexstrike-ai/`
- Domain: `cdn[.]discordapp[.]com`

Full set of 43 indicators unlocks with TL Creds: https://threadlinqs.com/tl-creds

## Sources

- [GTIG AI Threat Tracker: Distillation, Experimentation, and Integration of AI for Adversarial Use](https://cloud.google.com/blog/topics/threat-intelligence/distillation-experimentation-integration-ai-adversarial-use)
- [Google: China's APT31 used Gemini to plan cyberattacks against US orgs](https://www.theregister.com/2026/02/12/google_china_apt31_gemini/)
- [GTIG November 2025 AI Threat Findings](https://cloud.google.com/blog/topics/threat-intelligence/threat-actor-usage-of-ai-tools)
- [US DOJ Indictment of APT31 Members](https://www.justice.gov/opa/pr/seven-hackers-associated-chinese-government-charged)
- [OFAC Sanctions Against Wuhan XRZ](https://home.treasury.gov/news/press-releases/jy2205)
- [HexStrike AI Penetration Tool Citrix Exploits](https://www.theregister.com/2025/09/03/hexstrike_ai_citrix_exploits/)
- [Anthropic Report: Chinese Spies Abuse Claude Code AI](https://www.theregister.com/2025/11/13/chinese_spies_claude_attacks/)
- [Advancing Gemini Security Safeguards](https://deepmind.google/discover/blog/advancing-geminis-security-safeguards/)
- [Google Secure AI Framework (SAIF)](https://blog.google/technology/safety-security/introducing-googles-secure-ai-framework/)
- [Attackers experiment with Gemini AI Thinking Robot malware](https://www.theregister.com/2025/11/05/attackers_experiment_with_gemini_ai/)
- [APT42 Operations Analysis](https://cloud.google.com/blog/topics/threat-intelligence/untangling-iran-apt42-operations)
- [AI agents autonomous cyberattacks assessment](https://www.theregister.com/2026/02/03/autonomous_cyberattacks_not_real_yet/)

## 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-2026-0085
