decrypted · 17 november 2026 · ransomware and cybercrime · vulnerabilities and patching · ai and llm security

Google finds AI agents turning security patches into exploits

Every security patch is also a set of instructions. Compare the code before and after, and the difference points at the bug. This week Google's Threat Intelligence Group, GTIG, said it has seen that step automated for the first time.

What Google found

GTIG described an exposed staging server running a multi-agent framework it calls exploit_agent. It connects several frontier and locally hosted LLMs, the large language models behind chatbots, and is set up for what GTIG calls continuous differential analysis of popular web application software. It compares code before and after each patch announcement, works out what changed, drafts an exploit and runs it, with no human involved.

Picture a locksmith who posts a notice saying the faulty lock on a popular model has been fixed. A burglar once had to buy that lock and study it. Now a machine reads every notice, strips the old lock down and tries the result on every door it can find.

What the server held

Beside the framework sat stolen victim data and scripts for reconnaissance, password guessing and planting webshells (hidden back doors on a web server). GTIG said the victim telemetry suggested the operator was interested in French media, entertainment and education websites.

GTIG did not name the models, and its report does not say how often the generated exploits worked. Its worked example is access-control code in a WordPress plugin, Spectra, at version 2.19.25. GTIG also said such tools currently trade stealth for velocity, so they are noisy and easier to spot. Nothing names UK victims, but the targets are ordinary web software British organisations run too.

Why patched flaws are the target

GTIG suggested attackers may prefer patched flaws to zero-days, because a patch hands them a finite pile of source material, whereas hunting for unknown bugs means casting a wider net. The same report puts zero-day exploits at an average of 8 a month in 2025, 11 a month in 2026 and 22 in August. It also says researchers and Huntress used LLMs to produce working exploits for a PaperCut flaw within a day of disclosure, and that GreyNoise reported a campaign four days later that hit hundreds of organisations.

The Secure by Design lesson

Opinion, not advice: the patch window is now the product. If an exploit can exist within a day of a fix, a fix that waits for a monthly maintenance slot arrives late. Secure by Design puts that burden on the supplier. Automatic updates should be on by default, and administration screens should never face the open internet unless someone chooses to expose them.

For a UK organisation the cost is modest. Switch on automatic updates for the plugins and appliances you can, keep an inventory so you know what you are running, and measure the time between a patch appearing and it being installed. Then ask each supplier how fast they ship fixes and whether they push them to you.

Also this week

A chat server bug its maintainer calls the worst in 15 years. UnrealIRCd, open-source software for running internet chat servers, shipped version 6.2.7.1 and a hot-patch on 7 November. Its maintainer wrote on the oss-sec mailing list that an escaping bug in the IRCv3 label message tag lets ordinary users run commands with server or operator authority, and called it the most critical issue found in UnrealIRCd in 15 years. He said the nine security issues fixed in the release were found during an internal audit, with the help of AI. On 13 November CISA added CVE-2026-114713 to its KEV list, the catalogue of flaws known to be exploited in real attacks, citing evidence of active exploitation. The maintainer's own post does not say the flaw is being exploited. Few UK firms expose chat servers to the internet, but any that do should apply the hot-patch. Version 5 is end of life and affected too, so the maintainer's workaround there is to blacklist the labeled-response module and restart. Defenders using AI to audit their own code before attackers read the patch is the sensible answer to our lead story.

An employee data sale that began at a supplier. TSMC said the employee data leaked online matches data held by a cloud service provider, and that it found no evidence its own systems were compromised, according to BleepingComputer and TechRadar. A forum user with the alias SmithBianl posted a 499-row sample on 8 November and claims 46,000 employee records, roughly 7GB, including national ID numbers, addresses and banking details. Only the match to the provider's data has been confirmed. The size of the haul is the seller's claim, and the outlets disagree on what the asking price is worth in dollars, so treat the figure with care. TSMC has not named the provider. The Secure by Design lesson for UK organisations is about suppliers: your staff records are only as safe as the weakest company that holds a copy. Ask who stores your people data, why they keep it, and how quickly they must tell you when it goes missing.

Sources

If you want help checking how quickly your own systems get patched, get in touch.

More like this

Get the next post by email: subscribe to Decrypted. Double opt-in, unsubscribe any time, or take the RSS feed.

Prefer to listen? Decrypted on Apple Podcasts, or paste the podcast feed into any app.

Play here Loads Apple's player when you press it.