EXPLORINEER
en

VN — VnExpress

Google Detects First Fully Autonomous AI-Driven Cyberattack

Security2 min read

Google says it found the first real cyberattack that was run completely by AI, with no human steering it. The AI system found weak spots, wrote attack code, and hit websites in France.

What happened

Google said it found, almost by chance, an AI system that could find security holes, write attack code, and attack websites in France without human help. Google's threat intelligence team, GTIG, and the security firm Mandiant confirmed it. Older attacks used chatbots only to help write malware. This system did every step itself, from searching for weaknesses to using them.

Key facts

  • Google found a test server on the internet running a multi-agent system called exploit_agent.
  • The system was connected to several large language models (AI models that understand and write text) and ran locally.
  • The main targets were media, education, and entertainment websites in France.
  • Google did not blame any specific hacker group or country.

How the attack worked

The system used patch diffing. It compared code before and after a developer released a security fix, so it could spot the weakness that had just been repaired. It then wrote attack code and struck systems that had not yet installed the fix. After getting in, the server also held tools for scanning internal networks, guessing passwords automatically, and installing remote-control malware to keep access.

Teams of AI agents

Google says criminals are moving toward multi-agent systems. Instead of one tool, they split the job among several specialized AIs. One studies the target, another looks for holes, and a third writes and tests the attack code and decides the next step. Google listed open-source hacking tools that hackers are testing, including ARTEX, CAIRN, CyberStrikeAI, PentAGI, and HexStrike AI. One group changed PentAGI to scan many organizations in Asia from late August to early October.

Why it matters

AI-written malware still often has many errors, and automated attacks can leave traffic patterns that are easier to spot. Still, experts say the security landscape is changing fast. Google advises organizations to install updates as quickly as possible, protect cloud accounts closely, and soon use AI-based automatic defenses to keep up with faster attacks.

Glossary

GTIG
Google's threat intelligence team, which tracks cyber threats and hacker activity.
Mandiant
A cybersecurity firm that investigates attacks and helps companies defend against them.

Same story, different eyes

Ad

Add 150 words. Learn 10 a day.

Dicter word books send only the words you learn to your review notes.

SecurityArs Technica

How to tell a cyberattack from a plain old outage

A cluster of tech outages in the US over four business days, including Microsoft Exchange, FAA, JPMorgan Chase, United Airlines, and NSA networks, fueled speculation of cyberattacks. However, there is no evidence linking these failures or showing hackers caused them. Cybersecurity analysts say most digital breakdowns stem from the complexity of IT systems rather than malicious intent, as Cloudflare CEO Matthew Prince notes.

Securityheise online

BSI Raises Threat Assessment to "Tense to Critical"

Germany's Federal Office for Information Security (BSI) has rated the IT threat situation "tense to critical" for the first time since 2023, and the outlook as "critical." Vulnerabilities are now disclosed at a rate of about 160 per day, 35 percent more than in the previous reporting period, and the time between disclosure and first exploitation fell to about 8 days in Q2 2026. The BSI plans a new "Cyberindex" to measure threat changes.

SecurityThe Register

Cyberattack on service provider exposes payee data at 40 Irish credit unions

An attack on software provider Wellington IT exposed saved payee details, including names, IBANs, and sort codes, held by around 40 Irish credit unions. Minister Robert Troy criticized the company's delay in notifying the affected institutions, saying the breach occurred in October and most were only informed in recent days. Credit unions said login credentials and passwords were unaffected and the exposed data does not grant access to member accounts or funds.

SecurityThe Register

TfL takes Oyster photocard website offline after attackers compromise customer accounts

Transport for London (TfL) has taken its Oyster photocard website offline after attackers compromised customer accounts, potentially affecting around 265,000 customers. TfL says no passwords, bank or payment card details, or journey information were compromised, and it has notified the ICO and contacted those potentially affected. Customers will be asked to change their passwords and security questions, and existing photocards still work at ticket barriers while the site is down.