Google Warns Hackers Are Using AI to Build Working Zero-Day Exploits

Artificial intelligence has officially transitioned from an experimental hacking novelty into an industrial-scale weapon for cybercriminals.

Google Threat Intelligence Group (GTIG) adversaries are now actively using generative AI models to discover vulnerabilities and engineer functional zero-day exploits.

This marks a significant escalation in the cyber threat landscape, shifting AI’s role from a simple research assistant to a sophisticated engine for autonomous attack operations.

Google’s proactive threat hunting recently thwarted a major cybercrime campaign aiming to deploy an AI-generated zero-day exploit in a mass attack. The vulnerability targeted a popular open-source web administration tool by bypassing two-factor authentication (2FA).

Google Warns Hackers Are Using AI

Traditional security scanners look for memory corruption or syntax errors. However, frontier language models excel at finding high-level semantic logic flaws. The newly discovered 2FA bypass vulnerability relied on a hardcoded trust assumption by the developer.

While the code appeared functionally sound to traditional fuzzing tools, the AI model successfully contextualized the developer’s intent and identified the hidden logic gap.

According to a newly released May 2026 report confirmed that the exploit script heavily relied on AI assistance during its creation. The Python code contained distinct hallmarks of large language model (LLM) generation.

• The code included highly structured, textbook Pythonic formatting typical of LLM training data.
• The script featured extensive educational docstrings and detailed help menus.
• The threat actors accidentally left a hallucinated CVSS severity score generated by the AI inside the script.

State-sponsored hacking clusters linked to the People’s Republic of China (PRC) and the Democratic People’s Republic of Korea (DPRK) are heavily investing in AI-augmented vulnerability discovery.

These groups bypass standard AI safety filters by using expert persona prompts. For example, the threat group UNC2814 instructed models to act as senior C/C++ security auditors to evaluate extracted router file systems for remote code execution flaws.

Some groups are taking this a step further by training models on specialized datasets to improve their exploit reliability. Threat actors have integrated the “wooyun-legacy” database, a collection of over 85,000 real-world vulnerability cases, into code analysis plugins.

• APT45 automates thousands of repetitive prompts to analyze different CVEs and validate proof-of-concept exploits recursively.
• Adversaries use agentic frameworks like OpenClaw alongside intentionally vulnerable testing environments to refine payloads before deployment.
• Actors use temporal knowledge graphs to maintain a persistent state of the attack surface, allowing autonomous tools to pivot seamlessly based on internal reasoning.

Autonomous Malware and Evasion

AI is not just helping write malware; it is actively piloting it. The Android backdoor known as PROMPTSPY represents a massive leap toward autonomous attack orchestration.

By using the Gemini API, PROMPTSPY can independently read the device’s user interface, interpret the system state, and execute commands without human supervision.

The malware serializes the Android screen to XML and asks the AI model for spatial coordinates to simulate physical clicks and swipes. This allows the backdoor to capture biometric data, replay authentication patterns, and even render an invisible shield over the “Uninstall” button so victims cannot remove the app.

Beyond autonomous execution, suspected Russia-nexus groups are using AI to generate massive amounts of decoy code to hide their true intentions.

• The CANFAIL malware includes explanatory developer comments explicitly highlighting inactive filler content designed to confuse security analysts.
• The LONGSTREAM downloader uses seemingly benign, but inactive, administrative blocks, such as repeatedly querying the system’s daylight saving status, to look benign.
• The HONESTCUE script interacts with AI APIs to request specific obfuscation techniques for just-in-time self-modification.

AI tools are accelerating the scale of reconnaissance and social engineering campaigns. Adversaries use AI to instantly map out the internal hierarchies and third-party relationships of target enterprises.

This detailed intelligence allows attackers to craft highly specific phishing lures aimed at administrators or finance personnel, far surpassing the effectiveness of commodity bulk phishing.

In the realm of information operations, campaigns like the pro-Russia “Operation Overload” rely on AI to fabricate digital consensus. The group has deployed AI voice cloning technology to impersonate real journalists in synthetic media.

This allows threat actors to rapidly produce localized, persuasive deepfake content to push specific geopolitical narratives.

Threat groups such as UNC6780 are now targeting the software dependencies of AI environments as an initial access vector. These supply chain attacks aim to compromise machine learning models to pivot into broader networks.

By hijacking insecure integrated components, attackers can trigger disruptive downstream actions, such as extortion or ransomware deployment.

While hackers continue to innovate, defenders are utilizing the same technology to secure digital infrastructure.

Google is actively deactivating malicious API accounts and using AI agents like Big Sleep and CodeMender to automatically hunt down and patch software vulnerabilities before adversaries can exploit them.

Follow us on Google News, LinkedIn, and X to Get Instant Updates and Set GBH as a Preferred Source in Google.