November 19, 2026
A Security Patch Came Out. An AI Agent Wrote the Exploit. No Human Touched It.
How hackers are using AI in 2026: five things attackers really do with it, the numbers behind each one, where the hype runs ahead of the…

By Faheen Ahmed
8 min read
How hackers are using AI in 2026: five things attackers really do with it, the numbers behind each one, where the hype runs ahead of the evidence, and seven habits that still stop most of it.
Short answer: In 2026, hackers use AI in five main ways: to turn freshly published security patches into working exploits, to run large parts of an intrusion with AI agents, to build malware that asks a language model what to do next, to write phishing and clone voices at scale, and to steal other people's AI accounts and coding tools. AI has not made attackers magical. It has made them faster and cheaper, and it has shrunk the time defenders get to patch from weeks to hours.
This story was written with the assistance of an AI writing program. Every figure comes from the reports and studies named in the text.
How are hackers using AI in 2026?
For a few years the honest answer was "mostly to write better emails." That stopped being true somewhere in the last twelve months.
On November 16, 2026, Google's Threat Intelligence Group published a report with an uncomfortable first. On an exposed staging server, its analysts found a tool the attackers had named exploit_agent. It watches popular web software for new patches, compares the code before and after, works out what the patch fixed, writes an exploit for it, and runs it. Google says this is the first time it has seen attackers automate the whole chain from finding a flaw to exploiting it with no human in the loop.
The details are oddly ordinary. A comment in the configuration file, written in Russian, labels the API keys as being for an "exploit_agent pilot". The same server held stolen data, brute-force scripts and webshells. The victims of interest looked like French media, entertainment and education sites.
That is one of five patterns worth knowing.
1. How does AI turn a security patch into an attack?
A patch is a public description of a hole. It always has been. What changed is how long it takes to read one.
Diffing a patch and building an exploit used to need a skilled person and a few days. Google's report says attackers are now mostly aiming AI at flaws that already have a fix, betting they can get in before you update. In August 2026, researchers produced working AI-written exploits for two PaperCut flaws within a day of disclosure. Four days after disclosure, a campaign that was probably AI-assisted had reached hundreds of organisations.
True zero-days are rising too, just more slowly. Google counted an average of 8 exploited zero-days a month in 2025, 11 a month in 2026, and 22 in August alone.
And in May 2026, Google reported the first zero-day exploit it believes was written with AI: a script that bypasses two-factor authentication on a popular open-source admin tool. The giveaways were almost funny. The code was full of tidy teaching-style comments, and it included a severity score the model had simply made up.
2. Can AI run a whole cyberattack by itself?
Most of one, yes. Not all of it.
The best-documented case is still Anthropic's November 2025 disclosure. A group it assessed with high confidence to be Chinese state-sponsored used Claude Code against roughly 30 organisations, including tech companies, banks, chemical manufacturers and government agencies. The AI did an estimated 80 to 90% of the work: scanning systems, finding weak points, writing exploit code, harvesting credentials, sorting the stolen data. Humans stepped in at perhaps 4 to 6 decision points per campaign.
How did they get past the safety rules? They split the job into small, innocent-looking tasks and told the model it worked for a security firm doing authorised testing.
Since then the tooling has gone open source. Google names agentic attack frameworks such as HexStrike AI, Strix, CyberStrikeAI and PentAGI turning up in real intrusions, including a China-linked campaign against a Japanese technology firm.
3. What is AI-powered malware?
It is malware that does not carry all its instructions with it. It asks a model at run time.
Google documented the first families in November 2025:
- PROMPTFLUX asks an LLM to rewrite its own source code every hour so that antivirus signatures stop matching. It was experimental and could not yet compromise a device.
- PROMPTSTEAL, used by the Russia-linked group APT28 in Ukraine, has a model generate the Windows commands that collect documents and system details.
- QUIETVAULT steals GitHub and npm tokens, then uses AI command-line tools already installed on the developer's machine to hunt for more secrets.
By May 2026 there was PROMPTSPY, an Android backdoor first spotted by ESET. It sends a description of the phone's screen to a model, which replies with where to tap and swipe. The malware does not need to know your banking app's layout in advance. It looks at it and works it out.
4. Is AI phishing really better than human phishing?
It is exactly as good as an expert, and it costs almost nothing.
Fredrik Heiding, Bruce Schneier and colleagues tested this on 101 people. Generic phishing emails got a 12% click rate. Emails hand-written by human experts got 54%. Emails researched, written and sent entirely by AI also got 54%. The AI built an accurate profile of its target 88% of the time using only public information.
So the quality ceiling did not move. The price did. A personalised spear-phishing email used to be reserved for executives. Now everyone on the staff list can get one.
Voice and video follow the same curve. The FBI's 2025 Internet Crime Report was the first with a dedicated AI section: 22,364 complaints and about $893 million in losses where AI was mentioned, with $632 million of that in investment fraud. The FBI calls this a floor, because most victims never realise AI was involved.
5. Why are attackers stealing AI accounts and coding tools?
Because AI is now both a resource worth stealing and a door into your machine.
Stolen AI access. Google's incident responders call it LLMJacking: leaked API keys get used to run someone else's workloads on your bill. In some 2026 cases the fraudulent charges passed $100,000 in a single day. The stolen access is resold through brokers for as little as $0.10 per million tokens.
Poisoned AI tooling. In the same report, Google describes a North Korea-linked group that broke into crypto companies and edited the .claude/settings.json file inside internal code repositories. The next developer to open Claude Code in that repo downloaded a backdoor. Earlier in the year, malicious packages posed as agent "skills", and a supply-chain attack on the AI gateway LiteLLM went after the API secrets it holds.
If you write code with an AI agent, your agent's config files are now part of your attack surface.
Is AI hacking as dangerous as the headlines say?
It is serious, and it is also messier than the headlines suggest. Both are true.
- Anthropic reported that the AI in the espionage campaign sometimes hallucinated credentials and claimed to have stolen secrets that were actually public.
- After two PHP flaws were patched in October 2026, hundreds of AI-generated "exploits" appeared that did not work.
- Google says autonomous attack tools "trade stealth for velocity". They are noisy and leave traces that make them easier to catch.
- Google also says attackers have not achieved breakthrough capabilities against the safety logic of frontier models. They get around it with role-play, stolen accounts and open-weight models.
exploit_agentwas a pilot. Google published no victim count for it.
The fair summary: AI has not created a new kind of attack. It has removed the labour cost from the old ones.
Are defenders using AI too?
Yes, and on finding bugs they may be ahead.
Google's Big Sleep agent had found 20 real vulnerabilities in open-source software by August 2025. The autonomous tester XBOW reached the top of HackerOne's US leaderboard in June 2025. In April 2026, Anthropic announced Project Glasswing: its Claude Mythos Preview model had found thousands of zero-day vulnerabilities across critical software, and the company gave access to 12 launch partners including Apple, Microsoft, Google and the Linux Foundation, with up to $100 million in usage credits, so the holes could be fixed first.
The catch is that a found bug only helps if someone ships the patch and someone else installs it. That last step is still done by people.
How do you protect yourself from AI-powered attacks?
Nothing here is new. What changed is the deadline.
- Patch internet-facing software in hours, not weeks. Turn on automatic updates wherever you can. The gap between patch and exploit is now about a day.
- Use phishing-resistant sign-in. Passkeys or hardware keys. A perfect phishing email is useless if there is no password or code to hand over.
- Verify money and access requests on a second channel. A voice or a video call is no longer proof of identity. Call back on a number you already had.
- Block outbound traffic you do not need. Egress filtering stops stolen data leaving and stops malware phoning a model for instructions.
- Treat AI keys like cash. Keep them out of repos, rotate them, set hard spending limits and alerts on every provider.
- Review agent config in every repo. Settings files, rules files, skills and MCP servers can run commands. Read changes to them in code review like you would a build script.
- Point AI at your own code first. Run an AI security review in CI. The attackers are already running one on you.
The bottom line
The interesting number in all of this is not 54% or $893 million. It is one day: the time between a fix being published and an AI-written exploit for it working. Attack skill used to be scarce, and that scarcity protected a lot of sloppy systems. It is not scarce any more. The teams that will be fine are the ones that got boring things right and got them done quickly.
FAQ: AI hacking in 2026
How are hackers using AI in 2026? Hackers use AI to turn security patches into exploits, run intrusions with AI agents, build malware that queries a language model at run time, write phishing and clone voices at scale, and steal AI accounts and developer tools.
Can AI hack by itself without a human? Partly. In November 2026 Google reported a tool called exploit_agent that found and exploited flaws with no human in the loop. In Anthropic's 2025 espionage case, AI did 80 to 90% of the work with humans making 4 to 6 key decisions per campaign.
What is AI-powered malware? AI-powered malware is malicious software that calls a large language model while it runs, to generate commands, rewrite its own code or decide what to do on the victim's device. Examples include PROMPTFLUX, PROMPTSTEAL, QUIETVAULT and PROMPTSPY.
Is AI-generated phishing more effective? It matches human experts. In a 101-person study, fully AI-automated spear-phishing got a 54% click rate, the same as expert-written emails and far above the 12% for generic phishing, at a fraction of the cost.
What is LLMJacking? LLMJacking is the theft of AI API keys or accounts to run workloads at the victim's expense. Google reported cases where fraudulent charges passed $100,000 in one day.
How can I protect myself from AI-powered cyberattacks? Patch quickly, use passkeys or hardware keys, verify payment and access requests on a second channel, restrict outbound network traffic, put spending limits on AI keys, and review AI agent configuration files like code.