AI Aulis

AI NEWS IN ENGLISH · 17 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

← Issue 17 November 2026Paper

Google: attackers’ system exploited patched vulnerabilities with no human involved

AI-generated contentWritten by AI

cloud.google.com, siliconangle.com · 2 sources · Sources published: 16 November 2026

Google’s threat intelligence group says it has observed, for the first time, an attacker system that finds and exploits web application vulnerabilities with no human involved. Attackers are also stealing companies’ AI credentials and selling access to them, Google said.

Google Threat Intelligence Group (GTIG) and Mandiant described the finding in a report published on 16 November, which the news site SiliconANGLE has summarised.

Google said the system, called exploit_agent, was found on an attacker’s staging server that was openly accessible on the internet. The system splits work among several AI agents and was connected to both frontier models and locally run large language models.

The system compares the code of popular web applications before and after patch announcements. From the difference it infers a potential vulnerability, writes an exploit – a program that takes advantage of the flaw – and runs it automatically.

The server also held data taken from victims, along with scripts for tasks such as cracking credentials and installing web shells (remote-access programs left on a server). Based on the victim data, Google assesses that the attacker was particularly interested in French media, entertainment and education websites.

SiliconANGLE reported that GTIG’s previous report, two months ago, had not yet recorded fully autonomous attack chains against real targets. Google’s own report describes previously observed attacks as partly automated.

Patched vulnerabilities may be a more efficient target for attackers

GTIG assesses that attackers may find it more efficient to point large language models at already patched vulnerabilities than to hunt for zero-day vulnerabilities that have no fix yet. A patched flaw comes with a bounded set of material: code versions, release notes and research. When the models chain the steps together, the time from patch to attack can shrink considerably, Google said.

As an example, Google cites two vulnerabilities in PaperCut print management servers that were exploited as zero-days in late August. Researchers and the security company Huntress used large language models to produce working exploits for them within a day of disclosure. Four days after disclosure, GreyNoise observed a campaign against hundreds of organisations by an attacker who likely used large language models, Google’s report says. At one target, the attacker gained domain administrator privileges within minutes, GreyNoise said.

GTIG’s figures show that vulnerability disclosures and exploitation nearly doubled between January and August 2026. Exploited zero-day vulnerabilities averaged 8 a month in 2025 and 11 a month in 2026. In August there were 22.

Google said the open-source penetration testing tools attackers use currently trade stealth for speed. Network traffic, identifiable signatures and code traces make them easier to detect, but the tools are still effective at achieving their goals. Hundreds of non-working exploits, likely made with large language models, were also published for October’s PHP vulnerabilities, and checking them kept the security industry busy.

Stolen credentials used to run AI

Financially motivated attackers steal companies’ large language model credentials and cloud resources. In cases Mandiant investigated, unauthorised charges sometimes exceeded $100,000 a day and at worst totalled $1 million before remediation. Attackers resell access on LLM Mixer marketplaces, where advertised prices range from $0.10 to $25 per 1 million tokens.

Google said a suspected North Korean group modified .claude/settings.json, the configuration file of Anthropic’s coding agent Claude Code, in internal repositories of cryptocurrency organisations. After the change, Claude Code downloaded the SOMBERMEME backdoor on startup. Google suspects the MIDNIGHT NEPTUNE cluster, while SiliconANGLE names the group UNC6836.

Report an error in this story