cd /news/ai-safety/hackers-built-an-ai-that-turns-secur… · home › topics › ai-safety › article
[ARTICLE · art-170666] src=madrobot.blog ↗ pub= topic=ai-safety verified=true sentiment=↓ negative

Hackers built an AI that turns security patches into working attacks on its own, Google says

Google Threat Intelligence Group and Mandiant found the first observed AI system that converts software security patches into working exploits with no human in the loop, running on an exposed staging server and dubbed exploit_agent, Google reported Monday. The multi-agent framework plugs into multiple frontier and local LLMs, performs continuous differential analysis of popular web software such as Spectra WordPress plugin version 2.19.25, and automatically executes drafted exploits; Google said disclosures and exploits nearly doubled between January and August 2026, with zero-day exploits jumping from an average of 8 a month in 2025 to 11 a month this year and 22 in August. Mandiant also reported LLMjacking abuse, with single-day fraudulent AI charges in some cases topping $100,000 and the worst total reaching $1 million before being stopped.

by read4 min views5 publishedNov 16, 2026
Hackers built an AI that turns security patches into working attacks on its own, Google says
Image: Madrobot (auto-discovered)

The Google Cloud logo on Google’s stand at the Tech in Government conference in Canberra, October 2022 (illustrative). Image: Tom Worthington / Wikimedia Commons, CC BY 2.0, cropped

Hackers have been caught for the first time running an AI system that turns software security patches into working attacks with no human in the loop, Google’s threat intelligence team said in a report on Monday. Google says attackers are moving from typing prompts to running whole teams of AI agents, and warns that fully autonomous intrusions are getting closer.

How the patch-to-attack system worked #

Google Threat Intelligence Group (GTIG) and Mandiant found the system on an exposed staging server. Dubbed exploit_agent, it is a multi-agent framework that plugs into “multiple frontier and local LLMs”, though the report does not name them or the group behind it.

The framework runs continuous “differential analysis” of popular web software: it compares code before and after a patch is announced, works out what the patch fixed, drafts an exploit and “automatically executes them without human intervention”. One example in the report shows it inspecting access controls in version 2.19.25 of the Spectra WordPress plugin.

The server also held data stolen from victims and scripts for brute-forcing passwords, installing webshells and removing other attackers’ webshells. Its victim records point to French websites in media, entertainment and education. A configuration file written in Russian labels the setup “LLM API keys for auto-exploitation (exploit_agent pilot)”.

The gap between patch and attack is shrinking #

Google says AI is “shrinking the patch window for organizations”. In late August, attackers chained two PaperCut print-server flaws (CVE-2026-81578 and CVE-2026-82078). Researchers, including at Huntress, used LLMs to build working exploits within a day of disclosure. Four days after it, GreyNoise reported, a campaign that likely used LLMs hit hundreds of organisations, at one point going from a foothold to domain administrator within minutes.

Disclosures and exploits nearly doubled between January and August 2026, GTIG found. Zero-day exploits rose more slowly, from an average of 8 a month in 2025 to 11 a month this year, before jumping to 22 in August. The flood has a side effect: after PHP patched two AI-found flaws on October 21, hundreds of exploits that did not work, likely AI-written, were published, leaving defenders to check each one.

Several groups are now using open-source agent hacking frameworks in live intrusions, including ARTEX, the tool Microsoft tied to China-linked attacks last month, plus CAIRN, CyberStrikeAI and PentAGI. GTIG saw CAIRN used against targets including US aerospace firms, and a PentAGI-based platform scanning organisations across Asia from late August to early October. For now these tools “trade stealth for velocity”, the report says, but Google expects “reduced human involvement” and stealthier attacks as the models improve.

Stolen AI accounts sold through ‘LLM mixers’ #

Criminals are also breaking into companies’ cloud accounts to run up their AI bills, a practice known as LLMjacking. Mandiant says single-day fraudulent charges in some cases topped $100,000, and in the worst, total abuse reached $1 million before it was stopped. In one case a botnet made millions of AI API calls in about 24 hours on stolen credentials.

The access is then resold through what GTIG calls “LLM Mixers”, illicit storefronts that route buyers’ requests through a pool of stolen credentials and spread the traffic across victims to avoid detection. Adverts on Russian- and Chinese-language markets offer frontier and jailbroken models for $0.10 to $25 per million tokens; one vendor, APIMart, charges five times more once a prompt passes 100,000 tokens. It is a step up from the discounted stolen logins GTIG reported in September, and echoes the ‘Skynet mode’ agent Symantec found hijacking exposed models.

North Korean hackers turned Claude Code into a backdoor #

The report also describes a suspected North Korean cluster, MIDNIGHT NEPTUNE, targeting cryptocurrency firms in North America and Europe. It approached staff through fake recruiter profiles whose photos carried Google’s SynthID watermark, then, once inside, edited .claude/settings.json files in company code so that starting Claude Code downloaded a backdoor called SOMBERMEME. It is the same kind of trust in developer AI tools exploited in the hacked Cursor extension, and fits Microsoft’s picture of North Korea using AI at every step.

Google says the activity it saw on Gemini triggered its safety systems, and that it disabled the assets involved and used the cases to strengthen its classifiers and models. It has also released detection rules for ARTEX and points to the safeguards built into Gemini 4 Carbon, its strongest model, and its Fairwind Program, which gives vetted defenders access to advanced cyber capabilities.

Why it matters #

Patching has always been a race, and an AI that reads a fix and fires an exploit on its own makes that race far shorter for anyone running popular web software. The same tools are cheap and open source, so the time companies have to install updates is shrinking towards days.

── more in #ai-safety 4 stories · sorted by recency
── more on @google 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/hackers-built-an-ai-…] indexed:0 read:4min 2026-11-16 · —