Skip to content

AI

GTIG warns agentic AI automates vulnerability discovery and exploitation at machine speed

GTIG says financially motivated and state-sponsored groups have moved from experimenting with offensive AI frameworks to large-scale adoption.

https://breach-signal.com/artificial-intelligence/agentic-ai-automates-vulnerability-exploitation/

Share

Key facts

  • GTIG documented AI operationalization by DPRK, Iran, PRC, and Russia in late 2025; November 2026 reporting shows large-scale state-sponsored offensive framework adoption.
  • GTIG noted agentic AI interest for malware tooling in February 2026; by November 2026 actors weaponized ARTEX, CAIRN, CyberStrikeAI, HexStrike AI.
  • GTIG warns that threat actors use agentic AI workflows and offensive frameworks to carry out vulnerability discovery and exploitation at machine speed.
  • GTIG saw large-scale adoption of AI-based offensive frameworks by financially motivated and state-sponsored groups after experimental use.

GTIG, Google’s threat intelligence team, warned on Nov. 16, 2026 that financially motivated and state-sponsored groups are using agentic AI workflows — systems of autonomous AI agents — and offensive frameworks to discover and exploit vulnerabilities at machine speed against internet-exposed endpoints.

GTIG said attackers abuse open-source penetration-testing suites including ARTEX, CAIRN, CyberStrikeAI and HexStrike AI for fully automated attacks. The tools target internet-exposed endpoints with low-complexity flaws that large language models — AI systems trained to understand and generate text — can easily exploit. GTIG saw large-scale adoption of these AI-based offensive frameworks by financially motivated and state-sponsored groups after a period of experimental use.

The team found a framework called exploit_agent on an exposed attacker server. It integrates frontier and local large language models for multi-agent vulnerability research and weaponization. The server held stolen victim data, reconnaissance details, webshells and brute-forcing scripts, and the framework scans software updates on popular web applications for signs of flaws before attempting to exploit them automatically. Hackers already used it to fully automate the discovery and exploitation of flaws against French media, entertainment and education sites.

GTIG said the most recent quarter saw the first large language model framework attacks that fully automate vulnerability discovery and exploitation. Google said: “fully automate vulnerability discovery and exploitation without a human in the loop for the first time.”

Within three months, attackers can run unsophisticated operations with little to no human interaction, using multiple AI agents, GTIG estimates. The team projects that fully autonomous, sophisticated agentic attacks are likely within six months. Automated frameworks cut defender patching time and create false leads.

Hundreds of mostly AI-generated proof-of-concept exploits surfaced after PHP disclosed two flaws in PHP2Shell. The volume burdened defenders with verification and protection work.

GTIG links UNC6836 to Midnight Neptune, a group that targeted cryptocurrency firms in North America and Europe using modified Claude Code. Once inside, attackers altered Claude Code settings so that starting it would download the SOMBERMEME backdoor. Suspected North Korean hackers also used Claude Code to deliver malware.

Actors used CAIRN, a publicly available multi-agent penetration testing framework, against the U.S. aerospace sector. GTIG counts it among the ARTEX, CAIRN, CyberStrikeAI and HexStrike AI tools actors had weaponized by November 2026. The team had noted agentic AI interest for malware tooling as early as February 2026.

Last week, the FBI warned in an advisory that multiple threat actors had used ARTEX in attacks since August. The actors hit financial, government, technology, defense, healthcare and other sectors.

Google said state-backed groups, commercial vendors and political networks routinely use AI in influence operations. LLMjacking is widespread, with attackers profiting at significant cost to victim organizations.

The broader pattern predates the November warning. In a February 12, 2026 report, GTIG said it had tracked AI operationalization by DPRK, Iran, the PRC and Russia in late 2025, and that reporting from November 2026 shows large-scale state-sponsored adoption of offensive frameworks. The Google Cloud Blog report updates GTIG’s November 2025 findings on threat actor AI usage, and GTIG published the quarterly update after observing increased AI integration across the attack lifecycle in the final quarter of 2025.

GTIG observed in late 2025 that AI had been integrated across the attack lifecycle, boosting reconnaissance, social engineering and malware development. Despite that, the team said it had not seen advanced persistent threat or information operations actors achieve breakthrough capabilities that fundamentally change the threat landscape.

State-backed groups used Gemini for offensive tasks in 2025. APT31 used Gemini, adopting a cybersecurity persona, to automate vulnerability analysis and testing, and experimented with Hexstrike MCP tooling for code execution against U.S. targets.

UNC795, a PRC-based actor, turned to Gemini every day across the attack lifecycle for code troubleshooting, research and building intrusion capabilities. UNC795 activity triggered Gemini’s safety systems, which refused the actor’s attempts to create policy-violating capabilities. It built an AI-integrated code auditing capability, a signal GTIG read as agentic AI interest for intrusion support.

APT42 used Gemini for malware and offensive tooling development, covering debugging, code generation and exploitation research. APT41 used Gemini to speed up malicious tooling through knowledge synthesis, troubleshooting, code translation and explanations of open-source tool READMEs.

UNC2970, a North Korean government-backed actor, used Gemini to profile targets including cybersecurity and defense firms, job roles and salaries. APT42 also turned to generative AI such as Gemini to improve reconnaissance and social engineering through official e-mail searches and partner research.

UNC6418 gathered credentials and e-mail addresses using Gemini, then phished those accounts in a Ukraine and defense campaign. Temp.HEX, a PRC-based actor, used Gemini and other AI tools to profile individuals in Pakistan and research separatist organizations.

Information operations actors use Gemini for research, content creation and localization, but GTIG found no AI-generated information operations content in the wild. Separately, actors from China, Iran, Russia and Saudi Arabia created political satire and propaganda for online platforms and printed posters.

GTIG observed HONESTCUE samples in September 2025 using Gemini’s API to outsource functionality generation. It prompts the Gemini API to obtain C# code for second-stage malware that downloads and executes another payload. It compiles and runs the received C# code in memory through .NET CSharpCodeProvider, leaving no disk artifacts, and the actor hosted final payloads on Discord CDN.

HONESTCUE is the work of a single actor or small proof-of-concept group, GTIG suspects, citing a single VirusTotal submitter, linked it to no known cluster, and says its hard-coded prompt is not malicious on its own. The AI use aimed to enable multi-layered obfuscation that reduces the effectiveness of traditional network detection and static analysis. HONESTCUE’s stage-two code downloads and executes another payload, a technique GTIG contrasted with PROMPTFLUX’s just-in-time approach. GTIG said the maliciousness arises from how the prompt output fits into a larger malicious construct.

Frequent model extraction or distillation attacks from companies and researchers worldwide trying to clone proprietary logic were mitigated by GTIG. Model extraction speeds up AI model development at lower cost and amounts to intellectual property theft that violates Google’s terms of service. AI model providers should monitor API access for extraction or distillation patterns. During 2025, GTIG observed such attacks on Google’s AI models aimed at insights into underlying reasoning and chain-of-thought processes.

Knowledge distillation trains student models from teacher models by querying the teacher about domain problems and applying supervised fine-tuning. Distillation from Google’s Gemini models without permission violates Google’s terms of service, and Google continues to develop detection and mitigation techniques. GTIG said these attacks typically do not threaten average users, concentrating risk instead among model developers and service providers. The risk involves intellectual property and competitive advantage rather than service confidentiality, availability, or integrity for end users.

For more than a year, state-sponsored actors have repeatedly turned to Gemini for coding, scripting, research into targets, vulnerability research and post-compromise work, GTIG said. In the fourth quarter of 2025, GTIG spotted direct and indirect links connecting threat actor misuse of Gemini to real-world activity. GTIG said APT actors use Gemini to support reconnaissance and target development for initial compromise, a shift toward AI-augmented phishing enablement. GTIG described LLMs as a strategic force multiplier that lets actors rapidly synthesize open-source intelligence to profile high-value targets, identify defense-sector decision-makers and map hierarchies.

GTIG said adding AI tools to their workflows lets threat actors move from reconnaissance to active targeting faster and on a larger scale than human-paced attacks allow. GTIG expects threat actors to keep using AI throughout the attack lifecycle, covering malware creation, vulnerability research, reconnaissance and lure generation. GTIG frames its reporting as a way to arm defenders against the next phase of AI-enabled threats by identifying early indicators and offensive proofs of concept to help defenders proactively thwart malicious activity and strengthen classifiers and models.

After the actor used Gemini to gather credentials and e-mail addresses for a Ukraine and defense phishing campaign, GTIG disabled UNC6418 assets, taking direct action against the unattributed threat actor following its misuse of Gemini. For 2025, GTIG reported no direct frontier-model attacks launched by tracked advanced persistent threat or information operations actors. The team also saw late-2025 AI malware experiments without major shifts and expects more exploratory testing, including conventional capabilities such as the COINBAIT phishing kit.

Was this helpful?

Join the discussion

Please read our Comment Policy before commenting. First-time comments are held for moderation. By posting a comment you agree to our Terms of Use and Privacy Policy, which explain how your name, email address and comment are used.

Your email address will not be published. Required fields are marked *