AI website hacking examples
AI systems are now finding and exploiting real flaws in real software, and attackers have used them in live campaigns.
- On bug bounty platforms. In June 2025, XBOW reported that its fully autonomous penetration tester had climbed to the top of HackerOne’s US ranking, competing alongside thousands of human researchers and making nearly 1,060 vulnerability submissions. The findings included remote code execution, SQL injection, XML external entities, path traversal, server-side request forgery, cross-site scripting, cache poisoning and exposed secrets.
- In widely used code. In October 2024, Google Project Zero’s Big Sleep agent found an exploitable memory-safety flaw in SQLite before it appeared in an official release, which the team called the first public example of an AI agent finding such a flaw in widely used real-world software.
- In a competition on real projects. At DARPA’s AI Cyber Challenge final in August 2025, competing systems found 54 of 63 planted vulnerabilities, patched 43, and also discovered 18 real, previously unknown ones.
- In an attack. Anthropic reported in November 2025 that a state-sponsored group used its model to perform 80 to 90 percent of an espionage campaign against about thirty targets, with humans stepping in at only a handful of decision points. The attackers got past safeguards by splitting the work into small tasks that looked harmless on their own.
AI attack speed and scale
AI changes the speed and scale of website attacks more than the kinds of attack. AI tools mostly find the weaknesses human attackers always looked for. What changes is how many sites they can test and how quickly they act on a newly published flaw.
That speed was already rising before these tools spread. Google’s threat intelligence analysis found that the average time between a vulnerability’s disclosure and its exploitation fell from 32 days in 2021 and 2022 to five days in 2023. A site that patches monthly now spends most of each month exposed to anything published since the last update.
Website protection against AI attacks
Protecting a website from AI-driven attacks means closing the common doors, and closing them faster.
- Patch internet-facing software within days. Content management systems, plugins, frameworks and web servers first. CISA’s Known Exploited Vulnerabilities catalog lists flaws already being used in attacks; anything on it that you run comes first.
- Remove what does not need to be public. Staging sites, old admin panels, test APIs, debug endpoints and unused plugins are all attack surface. Put admin interfaces behind your identity provider or a VPN instead of the open internet.
- Use phishing-resistant sign-in for admins. CISA calls phishing-resistant MFA, such as FIDO and WebAuthn security keys or passkeys, the gold standard. Stolen passwords are cheap, and AI makes phishing pages more convincing.
- Keep secrets out of the front end and out of code. Any key in a browser bundle can be read. Store server keys in a secret manager and scope each to one job.
- Fix the usual flaw classes in your own code. Validate input, use parameterised queries, encode output and check authorisation on every request, since these are exactly what automated testing finds.
- Test yourself first. Run automated scanning, or an AI-assisted test, against your own site before someone else does.
Operational systems off the internet
Most of the systems that run a business do not need to be on the internet at all. The website has to be public, but the devices and tools behind your operations, such as handhelds, scanners, point-of-sale terminals, sensors, kiosks and the systems that coordinate them, rarely need to be reachable from the internet. Each one that is reachable is one more thing an automated attacker can find.
Offline Protocol is built to take that layer off the internet. Devices discover each other, verify identities with Ed25519 keys they generate themselves, message over MLS end-to-end encryption, keep shared state in sync as replicated documents and call each other’s services over Bluetooth LE, Wi-Fi and local IP, with no internet and no API key. Your backend stays the system of record and receives results when it is reachable. An AI scanning the internet can find your website; it cannot scan devices that coordinate locally and have no public address to find.