For years, the race between patch and exploit has had a human in the middle. A vendor fixes a flaw, attackers read the fix, work out what was wrong, write an attack and aim it at everyone who has not updated yet. That chain took days, and it took people. Google’s threat intelligence team says the people are now optional. In its quarterly AI Threat Tracker, published on November 16, the Google Threat Intelligence Group (GTIG) and Mandiant report that they have observed threat actors using large-language-model frameworks to fully automate vulnerability discovery and exploitation without a human in the loop — for the first time.

The report is the grimmest edition of the tracker yet, and it does not stop at one framework. The same document describes open-source agent frameworks running live intrusions, a suspected North Korean crew turning a coding assistant into a backdoor pipe, and a resale market for stolen AI access. The through-line is speed: every step of an attack that used to be manual is being chained together at machine pace.

The patch is now the map#

The centrepiece is a framework GTIG calls exploit_agent, found running on a staging server its operator left exposed to the internet. It wires together several frontier and locally hosted language models and runs what Google calls continuous differential analysis: it watches popular web software for security updates, compares the code before and after each patch, reasons about what the fix was fixing, drafts an exploit for the hole the update closed, and then runs it. No person approves any of those steps. The server also held scripts for post-compromise work — internal reconnaissance, brute-forcing, installing webshells, and removing other attackers’ webshells.

Google is careful about what it does not know. The report gives no success rate for exploit_agent, no count of sites it got into, and no name or country for whoever ran it. The only clue to origin sits in a configuration file, where comments in Russian ask for the language models’ API keys to be entered. The example Google chose to print is the framework inspecting access-control code in Spectra, a WordPress plugin at version 2.19.25. Google does not say the agent found a flaw in Spectra or exploited it; it shows the agent reading the code, and describes what the framework is built to do with what it reads.

Server racks in a dark data centre, one rack glowing red beneath a dissolving digital fingerprint
Editorial illustration: one rack burning red among the blue. AI-generated illustration.

The frameworks already in live intrusions#

exploit_agent is the headline, but it is not the only agent framework GTIG says it has seen in real attacks. The report names four open-source multi-agent penetration-testing tools that threat actors are abusing to automate phases of live intrusions, plus evidence of experimentation with custom-built harnesses. The pattern is the same in each case: agents discover and exploit low-complexity flaws, generate single-use payloads on demand, and adjust tactics from real-time execution feedback.

FrameworkWhat GTIG observed
ARTEXHighly publicised pentesting tool now abused in intrusions; Google has deployed detection rulepacks for it
CAIRNMulti-agent framework using an observe-orient-decide-act loop; seen in intrusion activity against targets including US aerospace
CyberStrikeAILLM-native framework wiring commercial and open-weight models to 100+ security tools for end-to-end attack chains
PentAGICustomised platform used for systematic reconnaissance and vulnerability scanning across Asia, late August to early October 2026
HexStrike AIMCP-integrated harness connecting LLM clients to 150+ offensive tools; experimentation observed

GTIG also reports evidence that a group with a suspected China nexus is developing its own automated attack framework for likely use in espionage. The barrier to entry, Google notes, is now low enough that entry-level operators can run complex campaigns — and the frameworks trade stealth for velocity: autonomous execution is forensically noisy and leaves recognisable identifiers and code-level traces. Google expects that trade to narrow as the tools and the models behind them improve.

Claude Code turned into a delivery pipe#

The most unsettling section of the report is not about exploits at all. It is about a coding assistant being turned against its own users. GTIG says a suspected North Korean cluster it tracks as MIDNIGHT NEPTUNE targeted cryptocurrency organisations in North America and Europe, starting with fake recruiter profiles on professional networking platforms — profile images that carried Google’s SynthID watermarks, confirming the identities were AI-generated. The personas directed targets to fraudulent meeting platforms running “ClickFix” lures, which talked victims into running supposed repair scripts.

Once inside, the group pivoted into the victims’ internal code repositories and modified .claude/settings.json files — the configuration that controls how Anthropic’s Claude Code behaves, including its execution hooks. From then on, every time a developer started Claude Code in an affected project, the altered hook triggered a download of a backdoor Google calls SOMBERMEME. To hide the lateral movement, the group created GitHub personal access tokens named after Microsoft Copilot, so the traffic looked like routine assistant activity. The attacker did not need to compromise Claude Code itself; it compromised the trust developers place in a checked-in configuration file.

A laptop screen showing glowing code with a padlock and gear, a red thread of light leading to a dark doorway
Editorial illustration: a settings file becomes a delivery pipe. AI-generated illustration.

Google published no hashes, domains or victim names for SOMBERMEME, and no incident dates. Anthropic’s response, if any, is not cited in the report. The practical takeaway is narrow and urgent: configuration files for coding assistants — settings, hooks, skills — are now part of the attack surface, and they deserve the same review discipline as the code they sit next to.

Stealing the models themselves#

The report’s third thread is about money. Throughout 2026, Mandiant responded to several “LLMJacking” incidents, where financially motivated actors broke into cloud environments and burned the victims’ compute and language-model quotas. The spikes can be enormous: in some cases a single day of fraudulent charges topped $100,000, and in the worst cases cumulative abuse reached $1 million before the account was remediated. In one investigation, a botnet or proxy service generated millions of LLM API calls in about 24 hours using credentials that had been sold or leaked on underground forums.

That theft now has a retail layer. GTIG is tracking “LLM Mixers”, illicit storefronts that sell unauthorised access to language-model services through pools of stolen credentials and compute, spreading query volume across victim environments to stay under the radar. Advertised prices run from $0.10 to $25 per million tokens, and the market sells access to frontier and jailbroken models alike, including models advertised as supporting malicious activity.

The numbers behind all of this are moving the wrong way. GTIG counted an average of 8 exploited zero-day vulnerabilities a month in 2025 and 11 a month in 2026 — with 22 in August alone — while overall vulnerability disclosures and exploits nearly doubled between January and August 2026. Two PaperCut print-server flaws exploited as zero-days in late August, CVE-2026-81578 and CVE-2026-82078, had working exploits generated with language models within a day of disclosure, according to the report; a campaign four days later hit hundreds of organisations. A flood of claimed-but-broken proof-of-concept exploits, likely generated with language models, now adds verification work for defenders on top of the real attacks.

What to watch#

Three things matter next. First, whether the patch window keeps shrinking: if exploit_agent-style pipelines become routine, the days between a vendor’s fix and a weaponised exploit could collapse to hours, and patching discipline becomes the whole defence. Second, whether the Claude Code vector spreads — other coding assistants share the same configuration-file pattern, and Google’s report is effectively a public warning to audit them. And third, whether defenders close the gap: the company points to its own countermeasures — detection rulepacks for ARTEX, a Mandiant hardening guide, the Big Sleep vulnerability-finding agent, CodeMender patching agents, and Gemini 4 Carbon, a model built with layered defences against cyber misuse. The report’s blunt message is that the offence has gone agentic first. The defence now has to catch up at the same speed.

Sources: Google Cloud Blog, “GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold”; SiliconANGLE, “Google says an AI framework turned security patches into exploits with no human in the loop”; Security Affairs, “Google GTIG Warns AI Agents Automate Exploits, Steal Credentials and Scale Cybercrime”; ExplainX, “SOMBERMEME Backdoor: DPRK Tampered Claude Code Settings”.