Confirmed 16 Nov 2026 · Security A one-minute read

Google found attack software that turns security patches into working exploits with no human.

Google Threat Intelligence Group published this on November 16, calling it the first time it has seen discovery and exploitation fully automated. The caveat: it names no operator and no victim count.

Until now

AI helps attackers. A person still finds the bug and fires the exploit.

Now

A patch can be read, weaponised and fired with nobody watching.

Why it matters

The patch is the tip-off. The framework watches popular web software, compares the code before and after an update, works out what was fixed, writes the exploit and runs it. Defenders have always had days of grace after a patch ships. That window is now whatever the loop takes.

Established

  • Google says it saw full automation with no human in the loop.
  • The framework ran continuous comparison of web software before and after patches.
  • The same server held stolen victim data and post-compromise scripts.

Not established

  • Who ran it. Google names no operator and no country.
  • How many victims. Telemetry only suggests French media and education sites.
  • That it beats a person. No success rate or yield is published.
The five-minute readThe record, the timeline, the weight and 3 sources

The record

Google Threat Intelligence Group published "Agentic threat activity crosses a significant threshold" on cloud.google.com on November 16, 2026. In its own words: "In the past quarter, GTIG identified threat actors using LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop for the first time." The evidence is an exposed staging server hosting what Google calls "a multi-agent vulnerability research and weaponization framework, dubbed exploit_agent, that integrates with multiple frontier and local LLMs". It "conducts continuous differential analysis of popular web application software. It compares code before and after patch announcements, analyzes findings to identify potential vulnerabilities, drafts exploits, and automatically executes them without human intervention." An environment file on the server carried comments for entering "LLM API keys for auto-exploitation (exploit_agent pilot)". The server also held stolen victim data and scripts for internal reconnaissance, brute-forcing, installing webshells, and removing other attackers' webshells. Victim telemetry suggested the operator was "specifically interested in targeting French websites associated with the media and entertainment industries, as well as the education sector". Google does not name the operator or a country; The Register, in a research article filed at 14:00 Universal Time on November 16, describes a suspected Russian-speaking operator and Russian-language tooling, which is reported and not in Google's text. The report's context figures: disclosures and exploits nearly doubled between January and August 2026; zero-day exploits averaged 8 per month in 2025, rose to 11 per month in 2026, and spiked to 22 in August 2026. Google also tracks other offensive frameworks: ARTEX, CyberStrikeAI, which connects models to more than 100 security tools, HexStrike AI, which wraps more than 150 tools behind more than 12 sub-agents, and PentAGI, used by one tracked actor from late August to early October 2026. On model abuse: single-day fraudulent charges above $100,000, cumulative abuse reaching $1 million in the worst cases, and one botnet generating millions of model API calls in about 24 hours. Google's page carries a 22:00 Universal Time metadata stamp, after this window closed; the earliest public reading we can fix is The Register at 14:00, inside it. Not established: which models it used, and whether any exploit succeeded.

How it got here

  1. Sep 18

    Google says one of its own models broke into three companies that had not agreed to be tested, 4 months earlier, and that the break-ins did not need disclosing. Editions 026 and 027.

  2. Nov 4

    OpenAI says its own pipeline built 948 working Linux root exploits, and a day later ships a hacking model it rates Critical to approved customers only. Editions 074 and 076.

  3. Nov 6

    Microsoft Threat Intelligence names a Cambodia-based criminal group running the ARTEX AI hacking framework against banks. Edition 081. A person was still driving it.

  4. Nov 12

    Palo Alto Networks says attackers used Chinese AI models to breach 10 organisations. Edition 089. The same day, researchers show a strong model delegating to a cheap sub-agent fails attacks it survives alone.

  5. Nov 16

    Google publishes the exploit_agent finding and calls fully automated discovery and exploitation a first, with the framework recovered from an exposed staging server rather than inferred from logs.

Who has to react

Every vendor that ships a security patch, because the patch is now the attack's starting gun; the frontier laboratories whose keys the framework was built to accept; the French media and education organisations in the telemetry; national cyber agencies that measure the patch window in days; and the insurers who price that window.

What to watch

  • Would strengthen itA national cyber agency, a CERT or another frontier laboratory publishes a dated finding of fully automated exploitation with no human approval step, naming a framework or campaign, by June 30, 2027.
  • Would strengthen itGoogle, Microsoft or a major vendor reports exploit_agent, or a framework built on it, used at scale against named organisations with a victim count, by March 31, 2027.
  • Would weaken itGoogle's next quarterly AI threat tracker records no further fully automated exploitation, or attributes the exposed server to security research rather than an operator, by March 31, 2027.

Why it weighs this much

Signal strength
CostMedium
CommitmentHigh
ReversibilityHard
PersistenceStructural

Cost is medium. Analyst work on a server Google found exposed, not a funded operation, but reading a recovered multi-agent framework and publishing a named first is checkable work. Commitment is high: Google published under its threat intelligence group's name, with quoted claims and a date, and a first claim is the kind of statement competitors test hardest. Reversibility is hard. A published report with a quoted finding cannot be withdrawn quietly, and the framework itself exists whatever Google says about it. Persistence is structural: once a patch can be read and weaponised without a person, the defenders' grace period after every future patch is shorter, and that does not reverse.

The words

Differential analysis
Comparing software before and after an update to find what was fixed, which points straight at the flaw the patch hides.
Webshell
A small script left on a hacked web server so the attacker can return and run commands later.
Zero-day
A flaw being exploited before a fix exists, so defenders have no patch to apply.
Human in the loop
A person who approves or triggers each step. Removing it means the software decides and acts alone.

Sources

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.

8 contributions. Thank you.

You have built
a record here.

Keep it: create an account to verify your contributions and put your name to the answers you choose. Your record stays yours.

Keep my record ↗
Optional: help show whose perspectives are missing
Kept with your answers under a random id, never with your name. Used for one thing: showing you, on your record, where you stand next to people like you, once at least 20 of them have answered.

Add your signature

Stand behind your judgment.

Create an account to attach your name to this answer. Your other answers show no name unless you sign them too.

Sign up to add your name ↗

Already have an account?