Google: AI agents now turn security patches into working exploits with no human in the loop
Google's threat team found an attacker framework that diffs new patches, writes exploits and runs them automatically. Patch internet-facing apps in days, not weeks.
Google's Threat Intelligence Group (GTIG) says it has seen, for the first time, attackers using AI agents to turn freshly published security patches into working exploits with no human in the loop. In practice, that means the gap between "a patch is out" and "attacks begin" is shrinking. Every organization with internet-facing web software should treat critical patches as days-not-weeks work, and teams using AI coding agents should review who can change the agents' shared configuration files.
What happened, in plain English
On November 16, 2026, Google published the fourth-quarter edition of its AI Threat Tracker, a regular report from GTIG and its incident-response arm, Mandiant. Google's summary puts it bluntly: attackers are now using "autonomous LLM frameworks" (LLM means large language model, the technology behind chatbots such as Gemini or ChatGPT) to watch for software updates, work out what each update fixes, and build an attack for it, without a person directing each step.
The clearest example GTIG describes is a toolkit it calls exploit_agent, found on an exposed server used by the attacker. According to the report as quoted by several outlets, it chains together several AI models, keeps comparing popular web software before and after each security update, works out which weakness the update closed, writes an exploit for it, and runs it against targets automatically. The same server held stolen victim data and scripts for the next steps inside a network: looking around, guessing passwords, and planting webshells (small hidden programs that give remote control of a web server).
An analogy: when a carmaker recalls a model to fix a faulty lock, the recall notice also tells thieves which cars have the weak lock. People have always read patches this way. What is new is that a machine now reads every notice, the moment it appears, and goes to try the lock itself.
GTIG did not name a group or a country. It describes the operator as suspected Russian-speaking, based on Russian-language tooling and notes, and says victim data pointed to French media, entertainment and education websites.
Are you affected? What to do now
Anyone who runs web software exposed to the internet is in the target pool. These tools are opportunistic: they look for whatever is unpatched, not for you in particular. Nothing in the report requires a new product, but some habits need to speed up.
Checklist
- Shorten patch deadlines for internet-facing systems. Measure from the date the vendor publishes its advisory, not from when your team noticed it. For critical flaws in public-facing web apps, aim for days.
- Check PHP now. GTIG cites two critical PHP flaws (CVE-2026-109618 and CVE-2026-109619) patched on October 21, 2026, where AI-assisted analysis and hundreds of claimed exploits followed quickly; many of those exploits did not work. The PHP project's fixed versions are 8.5.12, 8.4.27, 8.3.36 and 8.2.35. PHP's advisory says updating is the only complete fix, and that PHP 8.2 reaches end of life on December 31, 2026.
- Check PaperCut. GTIG also points to two PaperCut NG/MF print-management flaws (CVE-2026-81578 and CVE-2026-82078). PaperCut published its advisory on August 27, 2026. GreyNoise documented an AI-agent-run campaign from August 31 to September 2 that compromised 440 instances at 395 organizations in 48 countries; in the fastest case it went from first access to full control of the Windows domain in five minutes.
- Keep an up-to-date inventory of every internet-facing web application and its version. You cannot patch quickly what you do not know you run.
- Put a web application firewall (WAF) in front of exposed apps where you can. GreyNoise notes that a WAF stopped the PaperCut attackers on at least one target.
- Watch for webshells and fast lateral movement: new script files in web folders, unusual child processes spawned by web servers, and sudden credential dumping on servers. The report says these frameworks currently "trade stealth for velocity", which means they tend to leave noisy traces.
- If your developers use AI coding agents, treat the agents' shared configuration files (for Claude Code, the
.claude/folder committed to a repository) like code: require review for changes, for example with GitHub CODEOWNERS. Coverage of the report says one suspected North Korean group, which GTIG tracks as MIDNIGHT NEPTUNE, altered such a file in internal repositories so that starting the agent downloaded a backdoor. Claude Code's documentation confirms that shared project settings can define hooks (commands that run automatically) and that organization-managed settings override project files. - Lock down AI API keys. The report also covers "LLMjacking", where criminals steal cloud AI credentials and resell access. Keep keys in a secrets manager, rotate them, and set spending alerts and limits.
If you already patch internet-facing systems within days and review changes to developer tooling, your main task is to keep doing it.
The expert view
Patch diffing, comparing a binary or source tree before and after a fix to locate the bug, is decades old. It has always been the main way "n-days" (known, patched flaws) get weaponized. The skill bottleneck was the analyst: reading the diff, understanding the root cause, and turning it into a reliable trigger took experienced people hours to weeks. GTIG's finding is that this whole loop, from monitoring releases to diffing, reasoning about the root cause, drafting an exploit and running it against a target list, now runs as an orchestrated multi-agent pipeline with no human in the loop. That is the genuinely new part. The individual techniques are not.
It fits a pattern we have covered all autumn. GTIG names several open-source offensive agent frameworks in live use or testing, including CAIRN, PentAGI, ARTEX, CyberStrikeAI and HexStrike AI. Palo Alto Networks' Unit 42 reported CAIRN-driven espionage intrusions last week, and Microsoft, Amazon and the FBI have all published on ARTEX. The GreyNoise PaperCut analysis showed hundreds of agents (run on OpenAI's Codex harness with a DeepSeek model, according to GreyNoise) diffing installer versions to build exploits. Google's own Project Zero cited the same pressure when it announced shorter disclosure deadlines on November 16.
The numbers in the report support the "faster" story more than the "smarter" one. GTIG counted 22 exploited zero-days in August 2026, against an average of 11 a month earlier in 2026 and 8 a month in 2025. After the October PHP fixes, hundreds of purported exploits appeared within days, and most were likely AI-generated and non-functional. AI lowers the cost of trying, so defenders see more volume, more noise and a shorter window, but not necessarily higher quality per attempt.
Three practical consequences:
- The patch window is the attack surface. Exposure is now measured in hours to days after an advisory. Risk-based patching that leaves "medium" internet-facing issues for the monthly cycle deserves a second look.
- Behavioral detection beats signatures. Many indicators of these frameworks come from whichever model or tool they call, so static strings change easily. Their speed and noise (rapid scanning, many attempts in a row, quick credential theft) are easier to catch.
- Agent configuration is now supply chain. A repository file that tells an AI agent which commands to run automatically is executable policy. It deserves the same review as a CI pipeline definition.
What is still unknown: GTIG has not said how many victims exploit_agent reached or how reliable its exploits were, and attribution rests on language artifacts rather than a named group. We were unable to load the full GTIG report page at the time of writing, so some details above come from multiple news outlets quoting it; Google's own summary confirms the core finding.
Official sources
- Google: "The era of autonomous attacks has arrived" (GTIG AI Threat Tracker Q4 2026 summary)
- Google Threat Intelligence: Agentic threat activity crosses a significant threshold (full report)
- PHP.net: security releases 8.5.12, 8.4.27, 8.3.36 and 8.2.35 (October 21, 2026)
- GitHub advisory GHSA-w8hw-w9fx-j79w (CVE-2026-109618) and GHSA-6wmf-4rv5-p3rp (CVE-2026-109619)
- GreyNoise: AI-orchestrated global campaign against PaperCut NG/MF
- Claude Code documentation: settings files and precedence
- Coverage: The Register, CyberScoop, Security Affairs, The Hacker News