Just in
TechGoogle AI Threat Tracker Finds Exploits Run With No Human in the LoopTechJD Earnings: Q3 Net Income Rises 23% on 3.5% Revenue GrowthTechBank Outage Today: Chase, Bank of America, Wells Fargo Users HitTechBlackstone Amazon Data Center Deal: BXDC Takes 75% of 13 Sites for $5BTechNationwide Emergency Alert Test Set for Nov. 17 at 2:20 p.m. ETTechXPeng Earnings: Q3 Revenue Rises 19% to $3.62 Billion
Tech

Google AI Threat Tracker Finds Exploits Run With No Human in the Loop

Google's latest AI Threat Tracker says attackers have started using AI agents that find software flaws, write exploits and launch them on their own. The report also details stolen AI accounts and a backdoor planted through a coding assistant.

USA Publishing Group5 min read
Share
Silhouetted person facing several monitors full of code, illustrating AI agents that automate cyber attacks
Photo: Kevin Horvat / Unsplash

Key takeaways

  • Google Threat Intelligence Group said it found a tool called exploit_agent that compares software before and after patches, then writes and runs exploits without human help.
  • Zero day exploits averaged 11 a month in 2026, up from eight in 2025, before reaching 22 in August, according to the report.
  • Mandiant said some LLMJacking break-ins ran up more than $100,000 in fraudulent AI charges in one day, with the worst cases reaching $1 million.
In this story 6 sections
  1. Inside the exploit_agent server
  2. Open source attack kits in real intrusions
  3. Zero day counts jumped in August
  4. Stolen AI accounts and $1 million bills
  5. Coding assistants turned into a delivery route
  6. How Google responded

Google security researchers say they have, for the first time, caught attackers using an AI system that finds a software flaw, writes the exploit and launches it against targets without a person stepping in at any point. That discovery leads the latest Google AI Threat Tracker, a quarterly report from Google Threat Intelligence Group (GTIG) and Mandiant that Google, a unit of Alphabet (NASDAQ: GOOGL), published Monday, Nov. 16.

The broader message of the report is that hacking crews are moving past typing prompts into chatbots. Criminal gangs and state-backed groups are now handing entire stages of an intrusion to teams of cooperating AI agents, GTIG said. In a companion post, Google framed the problem bluntly: "Defending against machine-speed threats at human speed isn't feasible."

Inside the exploit_agent server

GTIG said it found an exposed staging server running a multi-agent tool it calls exploit_agent. According to the report, the tool plugs into several frontier and locally hosted language models and keeps comparing popular web application code from before and after each patch announcement. It then works out what was fixed, drafts an attack for the weakness and runs that attack on its own.

The same server held data stolen from victims, plus scripts for later steps such as mapping internal networks, guessing passwords, planting webshells and wiping out webshells left by competing attackers. Victim records pointed to French websites in media, entertainment and education, the report said. Comments in the tool's configuration file were written in Russian. The report does not name the operator or tie it to a government.

Open source attack kits in real intrusions

The report lists several freely available agent frameworks that it says have moved from labs into live attacks:

  • CAIRN, a pentesting framework where specialized agents share notes on a common board, was used against targets that included the US aerospace sector.
  • PentAGI formed the base of a customized platform that one actor used from late August to early October 2026 to scan and probe a wide range of organizations in Asia.
  • CyberStrikeAI connects language models to more than 100 security tools, and HexStrike AI gives AI clients direct access to more than 150. GTIG said it saw attackers experimenting with both.
  • ARTEX, a pentesting tool the report calls highly publicized, is now the subject of dedicated detection rule packs in Google Security Operations.

Users ranged from profit seekers to suspected espionage groups with ties to China, and one China-linked group appeared to be building its own automated framework, GTIG said. The researchers added one caveat: these tools favor speed over stealth and leave noisy forensic traces, which makes them easier to catch for now.

Zero day counts jumped in August

GTIG also tracked how many flaws were exploited each month. Overall disclosures and exploits nearly doubled between January and August 2026, the report said, while zero day exploits rose from an average of eight a month in 2025 to 11 a month in 2026, then hit 22 in August alone.

By our math, the 2026 monthly average is 37.5% above the 2025 level (11 compared with eight), and the August count is 2.75 times the 2025 average (22 divided by eight).

The report also points to a side effect: a flood of AI-written exploit code that claims to work but does not, which forces defenders to spend time testing and debunking it.

Stolen AI accounts and $1 million bills

A growing crime business GTIG calls LLMJacking involves breaking into a company's cloud account and running up its AI and computing bills. Mandiant handled several such cases in 2026, the report said. In some, fraudulent charges topped $100,000 in a single day, and in the worst cases abuse reached $1 million before it was stopped. In one incident, a botnet or proxy service fired off millions of AI API calls in about 24 hours before the stolen credentials were cut off.

Stolen access is then resold through what GTIG calls LLM Mixers, underground storefronts on Russian and Chinese language forums. Advertised prices ran from $0.10 to $25 per million tokens, according to the report.

Coding assistants turned into a delivery route

One of the more unusual cases involves a suspected North Korean cluster GTIG tracks as MIDNIGHT NEPTUNE, which went after cryptocurrency firms in North America and Europe. The group used fake recruiter profiles whose photos carried Google's SynthID watermark, a sign the images were AI generated, and lured targets into running "fix" scripts on bogus meeting sites. Once inside, it edited Claude Code settings files in internal code repositories so that opening the assistant would quietly download a backdoor called SOMBERMEME, the report said.

The report also says CALANQUE ION, the Iranian government-backed group formerly tracked as APT42, keeps using models including Gemini for reconnaissance and social engineering. Separately, a commercial contractor used the Gemini API while building an influence platform designed to run more than 100,000 social media accounts.

How Google responded

Google said the activity it found triggered Gemini's safety systems and that it disabled assets tied to the actors. Google DeepMind used the findings to strengthen its classifiers and models so they refuse similar requests, the company said. Google is also promoting a product it calls Google AI Threat Defense, which pairs Gemini and other frontier models with Wiz, its CodeMender patching agent and Mandiant staff.

The shrinking gap between a patch and a working exploit is the same pressure behind Google's recent move to shorten Project Zero's disclosure deadline, and it raises the stakes for fixes like the recent patch from Microsoft (NASDAQ: MSFT) for an exploited Outlook spoofing flaw. The full findings are in GTIG's AI Threat Tracker report, with a shorter summary on Google's blog. More security coverage is on our tech news page.

Google AI Threat Tracker: Autonomous Hacking Tools