What happened
In May 2026, specifically on May 11, Google Threat Intelligence Group (GTIG) published a report with a finding the security community had long expected: for the first time, Google identified a criminal group using a zero-day exploit that, with high confidence, was developed with the help of artificial intelligence.
The Python exploit let attackers bypass two-factor authentication (2FA) on a popular open-source, web-based system administration tool. Google did not name the tool. According to GTIG, the criminals were planning a mass exploitation event, but early discovery let Google alert the vendor, which fixed the flaw before the large-scale attack.
How Google concluded AI was involved
GTIG did not watch the model being used; the conclusion came from analyzing the code itself. The script had the typical fingerprints of LLM-generated code:
- an abundance of educational docstrings, explaining each step like a tutorial;
- a hallucinated CVSS score inside the comments;
- "textbook" Python formatting, with detailed help menus and a tidy ANSI color class.
Google said there is no evidence that Gemini was used, meaning the model behind the exploit came from another provider or was an open model.
Why the flaw matters
The vulnerability was not a classic memory or input-validation bug. It was a high-level logic flaw: the tool's developer had hardcoded a trust assumption, an exception that in practice broke 2FA. Traditional fuzzers and static scanners struggle with this kind of bug because the code appears to work correctly.
That is exactly where language models shine: they can "read the developer's intent" and notice the contradiction between the 2FA rule and the hardcoded exception. The exploit still required valid user credentials, but combined with leaked passwords that is a small hurdle.
The rest of the report
The same GTIG report covered other relevant points:
- PromptSpy, Android malware that uses the Gemini API to understand the victim's screen and act on its own, even blocking the uninstall button with an invisible overlay;
- groups linked to China and North Korea using AI for vulnerability research, including thousands of repetitive prompts to validate proofs of concept;
- agentic tools such as Hexstrike AI and Strix used in attacks against a Japanese technology company;
- a grey market of premium AI accounts automatically created and cancelled to dodge bans.
What this means for businesses
The time between discovering a flaw and exploiting it keeps shrinking. For small and medium businesses, the practical lessons are:
- Patch quickly internet-facing admin tools, even "small" open-source ones.
- Don't expose admin panels directly; use a VPN or zero-trust access.
- Monitor leaked credentials: the exploit needed a valid username and password.
- Treat 2FA as one layer, not an absolute guarantee.
Sources
- Google (official blog, 2026-05-11): https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/google-threat-intelligence-group-report/
- The Hacker News: https://thehackernews.com/2026/05/hackers-used-ai-to-develop-first-known.html
- SC Media: https://www.scworld.com/news/google-reports-first-known-ai-assisted-zero-day-exploit-in-the-wild
Want to know whether your exposed systems are ready for AI-accelerated attacks? KGA IT offers a free consultation: kga-it.com.