Hackers Used Claude, Grok, and Kimi to Automate a 400-Bank Attack
Hackers reportedly used Claude, Grok, and Kimi AI models to automate a cyberattack across 400 banks, a sign AI-driven hacking has arrived.
What happened with the 400-bank cyberattack
Reports surfaced that attackers used multiple AI models, including Claude, Grok, and Kimi, to automate a cyberattack that hit roughly 400 banks. Instead of relying on a single tool or a team of human operators working targets one by one, the attackers apparently stitched together different large language models to scale up reconnaissance, exploitation, or both across hundreds of institutions at once. The detail that matters most isn’t the exact headcount of victims. It’s the fact that multiple frontier AI models were reportedly chained together to do offensive security work at a speed and breadth no human red team could match alone.
This didn’t happen in isolation. It landed in the same stretch of time as a wave of newly disclosed vulnerabilities in PHP, the language that still powers a huge share of the web, including most WordPress sites. Security researchers and even the NSA pushed out urgent advisories telling organizations to prepare for emergency patches or take systems offline entirely. Separately, an Indian digital certificate authority, eMudhra, was found to have been using laughably weak private keys, in at least one case reportedly just the number four, for certificates meant to verify trust for Indian government bodies, state-owned banks, and other institutions. None of these stories are the same incident, but they share a common thread: AI tools are now good enough to find and exploit security flaws at a scale that’s forcing the entire industry to react in real time.
TL;DR
- Attackers reportedly used Claude, Grok, and Kimi together to automate a cyberattack spanning around 400 banks, suggesting multi-model orchestration is now a practical attack technique, not a hypothetical.
- The incident surfaced alongside separate, serious PHP vulnerabilities that prompted NSA advisories and emergency patch warnings for systems running WordPress and other PHP-based applications.
- Security researchers have been warning for months that AI-assisted offensive security would move from research demos to real-world attacks, and this cluster of events is being read as confirmation.
- A related story involves eMudhra, an Indian certificate authority whose weak private keys (including one reportedly set to the number four) undermined trust certificates used by banks and government bodies for roughly 18 months.
- Researchers have shown that AI models can generalize a single discovered exploit into a reusable “skill” that then gets applied to search for similar vulnerabilities elsewhere, compounding the speed of discovery.
- Large, well-resourced institutions and big tech companies appear comparatively better protected right now, while smaller banks, regional institutions, and unmaintained websites are the most exposed.
- Basic hygiene (updated software, multi-factor authentication, unique passwords, antivirus on every device) remains the most realistic defense available to individuals and smaller organizations today.
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
How did AI models get used to automate a bank attack?
The core shift is orchestration. Rather than a single hacker manually probing one bank’s systems, then the next, then the next, the workflow described involves multiple AI models working in combination, with different models potentially handling different stages of the attack: scanning for weaknesses, generating exploit code, or adapting an approach that worked on one target to try against another. Claude, Grok, and Kimi are all capable general-purpose and coding-oriented models, and in the hands of someone directing them toward offensive security tasks, they can generate and test attack paths far faster than a human team working alone.
This mirrors a pattern security researchers have already documented in legitimate vulnerability research. One widely discussed example involved researchers using AI models to dig into PHP’s memory-cleanup processes, the routines that run when a program runs out of memory and has to reclaim resources. That’s an area prone to subtle bugs, and the researchers found an “interruption point” in that cleanup process that could be chained into a remote code execution exploit against WordPress, meaning an attacker could run arbitrary code on a vulnerable server without ever logging in. Once a vulnerability like that was found, the researchers reportedly had the AI model generalize the method into a reusable skill, essentially a documented strategy for finding similar classes of bugs elsewhere. That skill then got applied again and again, each success refining it further.
That same mechanic, point-and-generalize, is what makes multi-model attacks on hundreds of banks plausible. Once an exploit pattern works against one bank’s infrastructure, an AI-driven process can try variations of it against the next 399 targets with minimal additional human effort.
Why is this different from past automated attacks?
Automated attacks aren’t new. Botnets, credential-stuffing scripts, and vulnerability scanners have existed for decades. What’s different now is that the automation layer includes models capable of reasoning through novel situations, writing custom exploit code, and adapting strategy on the fly, rather than just running a fixed script against a list of targets.
Security commentators have been predicting this shift for months, pointing to earlier AI security research projects (some run privately before wider release) as early warning signs. The skepticism at the time was that AI-assisted hacking was mostly a research curiosity or a marketing talking point for AI companies touting their models’ capabilities. The reported 400-bank incident, paired with the PHP disclosures and the eMudhra certificate failure, is being treated by many in the security community as confirmation that the capability has moved from theoretical to operational.
Is the PHP vulnerability connected to the bank hack?
Not directly, based on what’s publicly known. The PHP vulnerabilities and the eMudhra certificate authority failure are separate stories from the 400-bank AI-driven attack. But they’re part of the same broader moment: a stretch of time in which AI-assisted security research (some of it defensive, through initiatives aimed at patching critical software before attackers find the same flaws) and AI-assisted attacks are surfacing in parallel.
Seven tools to build an app. Or just Remy.
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
The PHP flaws matter because PHP underpins a large share of the internet, including most WordPress installations. Under the right conditions, the vulnerabilities reportedly allow code execution without authentication, meaning an attacker doesn’t need a login to compromise a server. The NSA’s advisory urged organizations to prepare for patches immediately and consider taking affected systems offline until fixes are applied. Researchers also noted that Cloudflare was already observing early attack attempts exploiting the flaw shortly after disclosure.
The eMudhra situation is a different kind of failure: a certificate authority, the mechanism browsers use to verify that a website is who it claims to be, used extremely weak keys for an extended period, reportedly around 18 months, undermining trust certificates relied on by Indian government bodies, state banks, and other institutions across South Asia and the Middle East.
Who is most at risk from AI-driven attacks like this?
Smaller and mid-sized institutions, along with regional banks outside the largest global players, appear to be the most exposed. Large American banks and major tech companies generally have more mature security operations, more frequent patching, and in some cases earlier access to AI-assisted defensive tools through security partnerships. Smaller banks, regional financial institutions, and organizations running unmaintained websites or outdated software are in a weaker position, partly because there’s often nobody actively monitoring for new vulnerability disclosures and applying patches quickly.
This creates an uncomfortable dynamic: the organizations with the fewest resources to defend themselves are also the ones most likely to be targeted by automated, AI-driven attacks, simply because they present the path of least resistance.
What can individuals do to protect themselves right now?
Security professionals recommending immediate steps point to a short list of practical measures. Run full antivirus and anti-malware software on every device, including phones. Keep browsers, operating systems, and apps updated, and prioritize browsers with strong security track records. Turn on ad blocking and disable autoplay of active content. Be cautious with links in text messages and emails, even from familiar-looking sources, since compromised legitimate websites can be used to attack visitors. Use two-factor or multi-factor authentication wherever it’s available, ideally through an authenticator app or hardware key rather than SMS alone. And use a unique password for every account, so a breach in one place doesn’t cascade into every other account you own.
For anyone running a website, the advice is straightforward but urgent: confirm that any relevant security patches have actually been applied, don’t assume automatic updates have run, and verify the fix is live rather than just scheduled.
Frequently Asked Questions
What AI models were reportedly used in the 400-bank cyberattack?
Reports indicate Claude, Grok, and Kimi were used together in an automated attack that targeted around 400 banks, suggesting attackers combined multiple AI models rather than relying on one.
Is this the same incident as the PHP vulnerability news?
No. The PHP vulnerabilities and the eMudhra certificate authority issue are separate stories that surfaced around the same time as the reported bank attack. All three reflect the same broader trend of AI accelerating both vulnerability discovery and exploitation.
How does AI make cyberattacks faster or more effective?
AI models can generate and test many potential attack approaches in parallel, generalize a successful exploit into a reusable method, and apply that method against new targets with minimal human effort, compressing work that once took human teams much longer.
Are big banks and tech companies safer from these attacks?
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
Based on current reporting, larger institutions and major tech companies appear better protected, largely due to more consistent patching, larger security teams, and in some cases early access to AI-assisted defensive research. Smaller and regional institutions appear more exposed.
What should I do if I think my information was exposed?
Update passwords immediately, especially on accounts using a password shared with other sites, enable multi-factor authentication everywhere it’s offered, and keep an eye on your financial accounts for unfamiliar activity in the days and weeks following a disclosed breach or vulnerability affecting services you use.