Google says it found an AI tool that wrote and ran exploits with no human involved
The same Google report says a suspected North Korean cluster it calls MIDNIGHT NEPTUNE used AI in intrusions at crypto firms in North America and Europe. Google does not say who ran the exploit tool.

Contents
Google's threat researchers said they found an AI-driven tool on an attacker's server that watched for software patches, worked out the flaws they fixed, and then wrote and ran exploits without a person involved. Google Threat Intelligence Group (GTIG) described the tool, which it calls "exploit_agent," in a report published on Monday, November 16, prepared with Google's Mandiant unit.
The same report says a suspected North Korean threat cluster that Google tracks as MIDNIGHT NEPTUNE used AI tools in intrusions at cryptocurrency organizations in North America and Europe.
What Google found on the staging server
GTIG said it identified an exposed staging server, a machine attackers use to store and prepare their tools, hosting a multi-agent framework connected to several commercial and locally run large language models (LLMs). According to the report, the framework continuously compares the code of popular web application software before and after patches are announced, looks for the weakness each patch fixed, drafts an exploit, and runs it automatically.
Google called this a first. In the past quarter, the report says, GTIG saw threat actors use LLM frameworks to fully automate finding and exploiting vulnerabilities "without a human in the loop for the first time."
The server also held data taken from victims and scripts for later stages of an attack, including internal reconnaissance, password brute-forcing, and installing webshells, which are hidden control pages planted on a hacked website. Victim records on the server suggest the operator was interested in French websites in the media, entertainment, and education sectors, according to Google. One figure in the report shows the tool inspecting access controls in a version of the Spectra WordPress plugin. The report does not say whether that plugin was attacked.
What Google does and does not say about who ran it
The report shows two comments, written in Russian, from a configuration file for the tool's API keys. In Google's English translation, they describe the setup as an "exploit_agent pilot" and tell the user to enter the keys in the file, "not in the chat."
Google does not attribute the tool to any group or country, and it does not say the operator is Russian. The report gives no confidence level for who ran the server, does not name the victims, and does not say how many sites were compromised. CoinDesk, which reported on the findings on November 18, likewise did not go beyond the language of the comments.
AI in intrusions at crypto firms
In the part of its report on North Korean state-backed actors, Google said that during 2026 a "suspected MIDNIGHT NEPTUNE threat cluster" combined AI-assisted social engineering with abuse of the AI tools inside target organizations. Google described this pattern in recent intrusions at crypto organizations:
- The attackers set up fake recruiter profiles on professional networking sites. The profile images carried SynthID watermarks, Google's marker for AI-generated images, which Google said confirms the identities were made with AI.
- Targets were sent to fake meeting sites that prompted them to run "fix" scripts, a lure known as ClickFix.
- Once inside, the attackers changed configuration files in internal code repositories so that starting the Claude Code AI coding assistant would download a backdoor Google calls SOMBERMEME.
- They named GitHub access tokens after Microsoft Copilot to make their activity look routine.
The report does not name the companies, say how many were affected, or say whether any funds were stolen. It does not link MIDNIGHT NEPTUNE to TraderTraitor or to the theft from Binance.
Binance's own account of its October breach, as reported by Fortune and SecurityWeek, also described fake job interviews and the abuse of an internal AI tool. On October 31, the FBI attributed that theft to North Korea's TraderTraitor, and North Korea rejected the accusation. Google's report does not connect the two.
The wider numbers in the report
GTIG said overall vulnerability disclosures and exploits nearly doubled between January and August 2026. Zero-day exploits, attacks on flaws before a fix is available, rose from an average of eight a month in 2025 to 11 a month in 2026, then reached 22 in August, according to the report.
Google also said it had seen open source AI attack frameworks such as CAIRN and PentAGI used against real targets. It added that such tools "trade stealth for velocity," leaving traces that make detection more likely. Google said it disabled assets tied to the activity it tracked and used the findings to strengthen its models' refusals.
What it could mean for crypto firms
In our assessment, the report describes two pressures that matter for exchanges, wallet providers, and other crypto companies. The first is time: if patches can be turned into working attacks automatically, the gap between a public fix and an attack on unpatched software narrows. The second is the developer workstation: Google's account of MIDNIGHT NEPTUNE, like Binance's account of its breach, centers on staff devices, fake hiring approaches, and the AI tools engineers use every day. Google has not published the scale of either activity, and the operator of the exploit tool has not been identified.
Sources
- Google Threat Intelligence Group — GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold (November 16, 2026)
- CoinDesk — Google finds self-running AI hacking tool with Russian comments in its files (November 18, 2026)
This article was researched and drafted with the help of AI tools, checked against its sources in a separate AI fact-check and edit, and approved for publication by Trading Metrics. Read our AI use policy.
This article is for information only and is not financial advice. Crypto assets and leveraged products are volatile and you can lose money. Do your own research and consider your circumstances before trading.