Google says it found an AI system turning patches into exploits unaided
Google's threat team says it has, for the first time, seen a fully automated system turn software patches into attacks. It also tracked a market for stolen AI access.
Key takeaways
- Google's threat team says it saw attackers fully automate finding and exploiting software flaws with no human involved, a first in its tracking.
- A system called exploit_agent compared software before and after patches, wrote exploits and ran them on its own.
- GTIG says open-source agent frameworks such as CAIRN and PentAGI have been used in real attacks.
- Stolen AI and cloud access is resold through 'LLM Mixer' services; some victims faced over $100,000 in fraudulent charges in a day.
- GTIG says these tools currently trade stealth for speed, which makes them easier to detect.
Attackers have begun using AI agents to find software vulnerabilities and exploit them with no person involved, Google Threat Intelligence Group (GTIG) says in its latest quarterly AI Threat Tracker, written with Mandiant and published on Nov. 17; some outlets reported it a day earlier. GTIG says it is the first time it has seen vulnerability discovery and exploitation fully automated in this way. Earlier editions of the tracker described only partly automated attacks, TechTarget notes.
What did Google find?
GTIG says it found an exposed staging server hosting a multi-agent system it calls exploit_agent, which works with several frontier and locally run AI models. According to the report, the system keeps comparing popular web applications before and after patches are released, looks for potential vulnerabilities in what each patch changed, writes an exploit and then runs it automatically. The server also held stolen victim data and scripts for later steps such as internal reconnaissance, password guessing and installing web shells.
Data on the server suggests the operator was interested in French websites in media, entertainment and education, GTIG says. Comments in a configuration file were written in Russian, according to the report and Security Affairs. Google does not name the operator.
Which tools are attackers using?
GTIG says attackers have moved from typing prompts into chatbots to running open-source, multi-agent hacking frameworks. It names ARTEX, CAIRN, CyberStrikeAI and PentAGI. GTIG says it has seen CAIRN and PentAGI used in real attacks, including CAIRN against targets in the US aerospace sector, and has seen experiments with CyberStrikeAI and HexStrike AI. It also says a group with suspected links to China appeared to be building its own automated attack framework, likely for espionage.
The report says the methods themselves are not new: the frameworks still rely on stolen credentials and unpatched systems, but they chain steps together much faster. As an example of how quickly patch windows are closing, GTIG points to two PaperCut print-server flaws that were exploited as zero-days in late August. GreyNoise reported a campaign four days after disclosure in which an attacker likely used AI models to build exploits and hit hundreds of organisations, the report says.
Other frameworks GTIG tracked automate only parts of intrusions. The Hill reports that Google put the overall threat at "Stage 5," in which AI carries out several attack stages with at least one human operator involved.
What else is in the report?
GTIG describes "LLMJacking," in which criminals steal AI and cloud credentials to run AI models on a victim's account and resell the access. Mandiant handled cases where fraudulent charges topped $100,000 in a single day, and the worst reached $1 million before the intrusion was cleaned up, according to the report. GTIG says it found underground "LLM Mixer" services advertising this access on Russian- and Chinese-language markets for $0.10 to $25 per million tokens.
The report also says state-backed groups linked to Russia, China and Iran, along with commercial influence vendors, now routinely use AI models in information operations. In intrusions at cryptocurrency organisations in North America and Europe, a suspected North Korea-linked cluster that GTIG tracks as MIDNIGHT NEPTUNE changed Claude Code settings files in internal code repositories so that starting the coding tool downloaded a backdoor, the report says.
Is there any good news for defenders?
For now, GTIG says these frameworks give up stealth for speed. Their heavy network traffic, recognisable identifiers and code left behind make them easier to detect, though the report says they remain highly capable of reaching their goals. GTIG expects them to become stealthier as the tools and underlying models improve. GTIG expects attackers who need to stay hidden, such as espionage groups, to be more cautious with such noisy tools, while many financially motivated criminals are likely to accept that trade-off.
Outside experts quoted by TechTarget said defenders will need to speed up. "Traditional patching cycles are becoming obsolete, and some fixes will need to happen in hours rather than days or weeks," said Jason Elrod, CISO at MultiCare Health System.
The report is Google's own research and also describes Google's security products, including Google AI Threat Defense, which Google says combines Gemini and other models with Wiz, CodeMender and Mandiant. Google says it has released detection rules for ARTEX and that Mandiant has published a hardening guide.
Disclosure: NextFish's AI drafting desk uses Claude, a model made by Anthropic.
Questions readers are asking
What did Google report?
Google Threat Intelligence Group says it found, for the first time, a system attackers used to find and exploit software flaws with no human in the loop. Other tools it tracked automate only parts of attacks.
What is LLMJacking?
Stealing AI and cloud credentials to run AI models at a victim's expense, often reselling the access. Google says some cases cost victims more than $100,000 in a single day.
What should organisations do?
Experts quoted by TechTarget say patching needs to speed up sharply. Google says it has released detection rules for the ARTEX framework and that Mandiant has published a hardening guide.
Sources
- Google Threat Intelligence Group: GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold
- TechTarget: Google finds AI agents autonomously turning patches into exploits
- Security Affairs: Google GTIG Warns AI Agents Automate Exploits, Steal Credentials and Scale Cybercrime
- The Hill: Google: 'The era of autonomous attacks has arrived'
- #threat-intelligence
- #ai-agents
- #cyberattacks
- #vulnerabilities
Get the day's top AI & tech stories on WhatsApp: Follow the NextFish channel
Get stories like this every morning.
Related
Cybersecurity2 min
Critical NGINX flaws can let attackers run code on web servers
F5 has patched five remote code execution bugs. Wiz says AI tools helped it find the two HTTP/2 bugs; no attacks in the wild have been reported.
Cybersecurity2 min
European Commission confirms it was hit by an AI-driven cyber campaign
Pay, leave and medical reimbursement tools were taken offline. The Commission says it has found no evidence of a data breach affecting staff as a whole.
Cybersecurity2 min
Ukraine and FBI arrest four suspects in alleged $40.5M ransomware ring
Prosecutors say the group posed as IT support, using AI-generated voice calls to trick staff into handing over access. At least 28 companies paid ransoms, they say.
Cybersecurity2 min
Palo Alto Networks to buy AI startup Applied Compute for $4.5 billion
The cybersecurity company is betting that businesses will run their own open-weight AI models and need them secured. Its CEO says the bet could outgrow its security business.