16 November 2026 · Google Threat Intelligence
Google reports the first attack in which an AI agent found and used a hole without a human
Google's threat intelligence team says it has observed an attack in which an AI agent found a vulnerability and exploited it from start to finish, with no human in the loop. The agent framework is called exploit_agent, the tooling is Russian-language, and the targets were in French media, entertainment and education. Google published the finding on 16 November and calls it a threshold, not an incident: the distance between finding a hole and using it has collapsed.
CapabilityConfirmedMore than one independent source, or a primary document.
What happened
The report is Google's own, published under its Threat Intelligence Group. It describes an intrusion chain in which the agent ingests a patch, works out what the patch fixes, locates the vulnerable code, writes an exploit, tests it and runs it. The humans are around the system, not inside the loop at the moment of exploitation. Google names the framework `exploit_agent`.
That is one case. The report's argument is the surrounding market. Google lists open-source agentic attack tools by name: ARTEX, CAIRN, CyberStrikeAI, HexStrike AI, PentAGI and Agent Zero. It describes rented access to models for offensive work under labels such as LLMJacking and LLM Mixers, at prices between 0.10 and 25 dollars per million tokens. It names a Chinese group, LEAKY CASTLE, and a North Korean one, MIDNIGHT NEPTUNE, which it says abused a `.claude/settings.json` file to drop a backdoor called SOMBERMEME.
The number that carries the claim is the pace. Google puts the average number of zero-days it tracks at eight a month in 2025, eleven a month in 2026, and 22 in August 2026 alone. Politico summarised the finding as the first fully automated AI cyberattack and attributed the activity to Russian speakers. Google does not say a state was behind it, and it does not name the victims.
How it workedtechnical detail
A patch is a map. When a vendor fixes a hole, the diff shows exactly which function was wrong and what the safe version looks like. A human researcher reads that diff and then writes code that reaches the vulnerable path before the fix arrives. The step that takes skill is turning the diff into a working exploit for a specific target. Google's claim is that an agent can now do that step alone, because the reasoning is text and the target is another piece of text.
That changes the economics in one direction. Defence still needs a person to decide what to patch and when. Attack now needs a model, a target list and a payment method. The report's price list, dollars per million tokens, is the clearest sign that this is becoming a purchase rather than a craft.
What it cost
Google names sectors, not victims: French media, entertainment and education. The report does not say what data was taken, whether the intrusions succeeded, or what the cost was. No named person was hurt in anything I read.
What may follow
This is the entry that makes the rest of the year read differently. If finding and using a hole is now a purchase, the defender's advantage in numbers disappears, and the bottleneck moves to how fast you can patch. Two weeks earlier, on 3 November, the British cyber agency published its annual review with the same conclusion and a date: an AI-powered ransomware attack on the UK's critical national infrastructure is \"highly likely\" by 2028. The launch came with money: 60 million pounds for a first phase of a national Cyber Shield and 70 million for a centre that finds vulnerabilities across critical infrastructure. The same week, four ordinary systems failed without any attacker at all. Put together, the lesson is uncomfortable: the systems are already fragile, and the tools for finding the cracks are getting cheaper.
What I do not know
I read Google's report and Politico's summary. I did not read the TechCrunch piece on the FBI and ARTEX, only its headline, so I have left the FBI's "under three hours" claim out of the account above. Google does not identify the victims, the state behind the Russian-language tooling, or the outcome of the intrusion. Whether the agent acted alone for the whole chain or only for the last step is the detail that matters most, and the public report does not settle it. From the NCSC I read the news release, not the full State of the Cyber Nation report behind it.
Read next
- Five days, four systems, no attack: how Microsoft, the FAA, Chase and United fell over
- The NSA's own classified network is down, and there is no evidence of an attack yet
- Anthropic opens a less-blocked Claude to security teams
- OpenAI gives NATO's cyber agency access to its models: up to 300 billion tokens a month
- A Chinese AI tool was used in attacks on South Korean banks
Sources
- Google Cloud: GTIG AI Threat Tracker — agentic threat activity crosses a significant thresholdprimary · main source
16 November 2026. The primary document. Source for exploit_agent, the named open-source tools, LLMJacking and LLM Mixers pricing, LEAKY CASTLE, MIDNIGHT NEPTUNE and SOMBERMEME, and the zero-day counts.
- Politico: Google says Russian speakers pulled off first-ever fully automated AI cyberattackpress
16 November 2026. The framing as a first, the Russian-language attribution, and the targets in French media, entertainment and education.
- TechCrunch: FBI says Artex AI open source hacking tool used to breach networks in 'under three hours'press · not read end to end yet
Zack Whittaker, 16 November 2026. I have only read the headline, so the FBI's timing claim is not used in the entry above.
- NCSC: Cyber chief — UK facing AI-powered attacks on critical infrastructure within two yearsprimary
3 November 2026. The State of the Cyber Nation review, the \"highly likely\" judgement for an AI-powered ransomware attack on UK critical national infrastructure by 2028, Richard Horne's \"tech on the cheap\" speech, and the 60 million pound Cyber Shield and 70 million pound NERV Centre announced alongside it.