Google says hackers now use AI agents to break into systems on their own
Google's threat team says attackers have moved from AI-written phishing to AI agents that run intrusions largely alone, shrinking the time defenders have to fix flaws.
The short version
- Google's GTIG and Mandiant say attackers have crossed "a significant threshold" by using autonomous AI agents.
- Named open-source tools include ARTEX, CAIRN, CyberStrikeAI and PentAGI, which chain several AI agents together.
- AI can compare software before and after a patch and build a working exploit, sometimes within hours.
- Criminals also resell stolen access to AI models, a practice the report calls LLMjacking.
- One commercial platform was built to manage more than 100,000 fake social media accounts.
Why it matters
The tools are free and need less skill than traditional hacking, so more people can attempt attacks and each attempt costs less. For companies, the time between a security fix and a working attack is shrinking, so updating promptly matters more. For everyone else, the old basics still apply: two-step login and careful clicking.
What Google says has changed
As The Pioneer on MSN reported, Google's Threat Intelligence Group (GTIG) and Mandiant, its incident-response arm, say cyberattacks have passed "a significant threshold." Mandiant is the team Google uses to investigate break-ins. The finding comes from GTIG's quarterly AI threat tracker, published on 16 November 2026.
Until recently, criminals mostly used chatbots for chores like drafting phishing emails, the fake messages meant to trick people into handing over passwords. A human still had to steer each step. The report says that is no longer the whole picture.
Attackers are now running AI agents, programs that take actions on their own instead of only answering questions. These agents hunt through networks, look for weak spots and get inside, with little or no person involved. The report describes this as happening at "machine speed."
How the new attack tools work
The report points to multi-agent frameworks, which are bundles of several specialised AI agents working as a team. Named examples are ARTEX, CAIRN, CyberStrikeAI and PentAGI. All are open-source, meaning anyone can download and modify them for free.
One agent might probe a target, another picks out a weakness, and another writes a custom attack script on the spot. The team then changes its approach depending on what it finds. A single operator can start the process and let it run.
Google says it has seen this in the wild. CAIRN was used against a company in the US aerospace industry. A group linked to China used a modified PentAGI to probe many organisations across Asia from late August to early October 2026.
The low barrier is why the report matters beyond experts. Free tools that need less know-how widen the pool of people who can attempt a serious intrusion.
Patch day becomes a race
The report's most concrete finding concerns n-day vulnerabilities. These are security holes the software maker has already fixed, but which many organisations have not yet installed the fix for. The fix itself is called a patch.
Google found attackers using AI to compare a program's code from before and after a patch. The AI works out what was broken, then builds an exploit, a working method for abusing the flaw. Sometimes this happens within hours of the patch being released.
Google documented one case of a fully autonomous tool with no human in the loop. It kept watching popular web software for fresh patches, wrote exploits, and entered victim systems by itself. The operator appeared to be aiming at media and education websites in France.
Two other episodes show the speed. When two flaws in PaperCut, a print-management product, were being exploited in August 2026, both researchers and attackers had working exploits within a day of disclosure. After two PHP flaws were disclosed in October 2026, hundreds of bogus exploit claims, built with AI help, spread online within days. Many were fakes that did not work, which adds noise for defenders.
Stolen AI access and fake crowds
Two other trends stand out. The first is LLMjacking: criminals steal a company's credentials for AI services and cloud computing, then sell access to the stolen AI models through underground services the report calls "LLM Mixer" operations. Any company running AI work in the cloud could be a target.
The second is influence operations. The report says state-linked and commercial groups now use AI to run whole networks of fake personas, not just to write posts. One commercial operator was building a platform to manage more than 100,000 fake accounts.
The personas imitate regional dialects and slang. The platform adds deliberately imperfect, smartphone-style photos to slip past tools that spot AI-made images, and it automatically "ages" accounts so they look established before use. The report names campaigns aimed at Europe and the Middle East. It does not name one aimed at India, but notes that this kind of commercial tooling is sold to whoever pays.
What defenders and everyday users can do
Google says defenders are using AI too. It points to detection rulepacks, a Mandiant hardening guide, and tools called Big Sleep and CodeMender that use AI to find and fix flaws before attackers do.
For IT and security teams, the report's message is narrow: install security updates faster. A grace period of days or weeks is a poor fit when exploits can appear within hours.
For individuals, the report gives no attack counts, so this is a reason to cover the basics, not to panic. Automation tends to mean more attempts for less effort, and the weak spots are still the familiar ones: reused passwords and unexpected links.
What to watch
- Will Google or other security firms publish counts showing how common agent-driven attacks are?
- Will software makers and companies shorten how quickly they roll out patches?
- Will the report's authors or others document agent-driven campaigns aimed at India?
Try it
Turn on two-step login (a code sent to your phone) for your email and bank accounts today, and use a different password for each.
Sources and links
First reported by The Pioneer on MSN. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.