Technology

Cybersecurity News: North Korea Backdoors Claude Code

Cybersecurity news took a dark turn on Monday, November 16, 2026, when Google's Threat Intelligence Group revealed that a suspected North Korean hacking crew had turned one of the most popular AI coding assistants into a delivery system for malware. The group, which Google tracks as MIDNIGHT NEPTUNE, broke into the internal code repositories of cryptocurrency companies in North America and Europe and quietly edited a single configuration file, .claude/settings.json, so that every time a developer started Anthropic's Claude Code, the tool automatically downloaded a backdoor that Google calls SOMBERMEME. The finding was published in Google's quarterly AI Threat Tracker, a report that also documents, for the first time, attackers using an AI framework to turn security patches into working exploits with no human in the loop. Together, the two disclosures describe a new phase of cyber conflict in which the tools built to help developers write software faster are being turned against them, and in which the speed of attack has collapsed from days to hours.

The Google report, titled "Agentic Threat Activity Crosses a Significant Threshold," was published on November 16 by the Google Threat Intelligence Group, known as GTIG, together with the security firm Mandiant. It is the grimmest edition of the quarterly tracker yet, according to the coverage that followed, because it moves beyond warning about what AI-powered attackers might do and documents what they are already doing in live intrusions. The report describes open-source agent frameworks running real attacks, a suspected North Korean crew turning a coding assistant into a backdoor pipe, and a resale market for stolen access to powerful AI models. Google said it has disabled the accounts and assets tied to the misuse of its Gemini models that it found, and that Google DeepMind has used the findings to strengthen its classifiers. The report does not name the victims of the Claude Code attack, and it publishes no hashes, domains or incident dates for the SOMBERMEME backdoor, which means defenders know the technique but not yet the specific fingerprints to look for.

The attack on Claude Code began, as so many North Korean operations do, with a fake job. According to Google, the MIDNIGHT NEPTUNE cluster approached employees at cryptocurrency organisations using fake recruiter profiles on professional networking platforms, with profile images that carried Google's SynthID watermarks, confirming the identities were AI-generated. The personas directed their targets to fraudulent meeting platforms running what security researchers call ClickFix lures, which talk victims into running a supposed repair script on their own machines. That first step gave the attackers a foothold inside the companies. What happened next is what makes this case different from a routine intrusion. Once inside, the group moved into the victims' internal code repositories and modified .claude/settings.json files, the configuration that controls how 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 the SOMBERMEME backdoor onto that developer's machine.

Advertisement

How the Claude Code Backdoor Worked

The elegance of the attack, from the attacker's point of view, is that it did not require compromising Claude Code itself. The tool was working exactly as designed: it read a configuration file from the project it was asked to work on, and it ran the hooks that file specified. The problem was that the configuration file could no longer be trusted, because the attackers had edited it after gaining access to the repository. Developers place a particular kind of trust in checked-in configuration files, treating them as part of the project rather than as code that could be hostile, and the attackers exploited exactly that trust. To hide their lateral movement through the victims' networks, the group created GitHub personal access tokens named after Microsoft Copilot, so that the traffic generated by their activity looked like routine traffic from an AI assistant. The attacker did not need to break the tool. It only needed to poison the instructions the tool was given.

Google's report is careful about what it confirms and what it does not. It confirms the technique, the cluster label and the target profile: a suspected North Korean group, tracked as MIDNIGHT NEPTUNE, targeting cryptocurrency firms in North America and Europe. SiliconANGLE, reporting on the same document, also gives the group the alternate designation UNC6836. What Google did not publish is any indicator of compromise for SOMBERMEME: no file hashes, no command-and-control domains, no victim names and no incident dates. That absence matters for defenders, because it means the only way to check for this attack is to audit the configuration files themselves, looking for hooks or commands in .claude/settings.json that nobody on the team remembers adding. The report also notes that the group has evolved beyond the resume-generation and fake-interview tactics it was known for earlier in 2026, and that the Claude Code attack represents a supply-chain pivot against coding agents specifically, rather than a one-off trick.

The same report documents a second development that may prove even more consequential. GTIG says it has observed, for the first time, threat actors using a large-language-model framework to fully automate vulnerability discovery and exploitation without a human in the loop. The framework is called exploit_agent, and it was found running on a staging server that its operator had left exposed to the internet, wired into several frontier and locally hosted large language models. According to Google, the framework watches for newly published security patches, works out what vulnerability each patch fixes, and builds a working exploit before the patch has been widely applied. In the previous edition of the quarterly tracker, published two months earlier, GTIG said it had not observed fully autonomous attack pipelines deployed against targets in the wild. That changed over the past quarter. The report gives no success rate for exploit_agent, no count of the sites it got into, and no name or country for whoever ran it, but the fact that such a system was found operating at all is the point: the loop from patch to exploit, which used to require a skilled human analyst, can now run on its own.

Advertisement

The FBI Warning on AI Hacking Tools

Google's report did not arrive in isolation. Three days earlier, on Friday, November 13, the FBI issued a FLASH cybersecurity advisory warning that hackers are using an open-source AI penetration testing framework called ARTEX to automate cyberattacks against organisations in the United States and around the world. The bureau's alert, which consolidates 11 earlier reports rather than describing a single new intrusion, says ARTEX "significantly reduces the time required to chain attack techniques, with some intrusions progressing from initial targeting to privileged access in under three hours." The warning was issued alongside a separate alert from Amazon Threat Intelligence, and the two together describe criminal groups that have turned autonomous AI agents into a tool for running break-ins against banks, government agencies, energy suppliers and hospitals at what Amazon calls industrial scale.

ARTEX is an autonomous penetration testing framework written in the Go programming language. It connects large language models to software agents that turn plain-language instructions into commands, performing reconnaissance, discovering vulnerabilities, exploiting them, escalating privileges and stealing data with little human intervention. It was originally released on GitHub in July 2026 by a developer known as Autumn-27, who pulled the repository on October 8 after reports emerged that threat actors were abusing it. But by then the code was already in circulation, and the FBI says multiple threat actors have weaponised it since at least August to target organisations across the financial, government, technology, defence, healthcare and other sectors. The internet scanning service Censys counted 1,465 hosts running an exposed ARTEX console as of November 4, and the FBI said the true number of deployments is likely far higher once systems behind firewalls and on private networks are counted. Other researchers put the number of exposed consoles above 1,600 by early November.

The FBI's advisory cites two operations that show what ARTEX makes possible. In August, attackers used the framework against South Korean financial institutions, in an operation blamed in large part on weak cybersecurity defences and missing protections such as multi-factor authentication, stealing personal data belonging to roughly 67,000 people. In September, a supply-chain compromise of the IT company ConnectWise's Cloud Backup service ultimately affected more than 2,250 downstream Microsoft tenants, giving the attackers access to their Microsoft 365 cloud environments. Microsoft had previously linked ARTEX to Storm-0940, a China-state-aligned group targeting Microsoft 365 environments through compromised ConnectWise application credentials. Amazon's own honeypot network, known as MadPot, recorded two distinct ARTEX campaigns between October 3 and November 5: one that scanned exposed AI applications, stole provider keys and pooled them so the keys could be used to run models that searched for more keys, and a second that generated deepfake videos to trick live selfie identity checks. Amazon said it holds moderate confidence that the agents ran on ARTEX and moderate confidence that the operators were separate China-linked groups, citing simplified Chinese prompts and China time-zone settings, while cautioning that it did not watch the activity on the targeted systems as it happened and cannot verify the outcomes the agents reported.

Advertisement

What Google Found Beyond the Backdoor

The GTIG report covers more than the Claude Code attack and the autonomous exploit framework. It also describes a trend Google calls LLMjacking, in which threat actors steal unauthorised access to high-performance AI models and resell it through underground proxy services, letting other criminals run powerful models without paying for them. And it documents a specific software weakness, tracked as CVE-2026-40217, in the LiteLLM gateway, which allowed an attacker to turn a limited inference credential into full administrative access. The targets named in the captured records include model gateways and proxies such as One API, New API, Veloera, Sub2API and LiteLLM, alongside application builders including Dify, FastGPT, Langflow and RAGFlow. The picture that emerges is of an underground economy built on top of the AI boom itself: stolen model access, automated exploitation, and poisoned developer tools, all feeding each other.

The report also underlines a shift in how attacks are timed. Google observed individual automated bursts lasting two to 25 seconds, leaving defenders almost no time to assess connected events before the next wave arrived. The FBI made the same point in its own way, noting that ARTEX-linked attacks can be noisy, generating reams of traffic and data that may trigger a targeted company's defensive alerts, but that the speed of the attack can outrun the response even when the alarms do go off. The bureau's three-hour figure should not be read as a universal deadline for every cyberattack; it describes the speed observed in some ARTEX-enabled intrusions, not an average across all attacks. But the direction of travel is clear. The gap between the moment an attacker first probes a system and the moment they hold privileged access inside it is shrinking, and the tools doing the shrinking are the same AI systems that companies are racing to adopt.

For defenders, the practical lessons of the two reports are specific. Google's guidance on the Claude Code attack is to treat checked-in configuration files as untrusted input, auditing the hooks and commands in files such as .claude/settings.json with the same suspicion applied to any other code, and to verify that settings and hooks in agent configuration files have not been changed by anyone outside the team. The FBI urged organisations, particularly those running internet-facing applications, self-service portals, cloud infrastructure and supplier systems, to shrink their external attack surface, confine sensitive services to trusted networks or dedicated administrative paths, restrict administrative access, monitor changes to privileged identities, and preserve forensic evidence without delaying urgent containment. The bureau also asked organisations reporting ARTEX-related incidents to provide the time from initial suspicious activity to detection, and from detection to containment, so that it can measure how fast these attacks are actually moving.

Advertisement

What Happens Next for AI Security

The two reports, taken together, mark the moment when AI-powered cyberattacks stopped being a prediction and became a documented fact of life. In the previous edition of its tracker, Google said it had not observed fully autonomous attack pipelines in the wild. Two months later, it found one. The FBI, in the same week, warned that the tools for automating attacks are now available to actors with minimal technical expertise, enabling complex, multi-target operations to run simultaneously. The combination means that the advantage in cyber conflict is shifting toward whoever can automate fastest, and that the organisations most at risk are not only the ones with the weakest defences but the ones whose developers rely most heavily on AI assistants and agent frameworks. A coding assistant that reads its instructions from a shared repository is only as trustworthy as that repository, and a patch that is published on Monday can be turned into an exploit by Tuesday, with no human involved at any stage.

There are limits to what the public reports can tell defenders. Google published no indicators of compromise for SOMBERMEME, so there is no hash or domain for a security team to block. The FBI cautioned that the IP addresses listed in its advisory may already have changed. And neither report can say how many organisations were affected by the Claude Code attack, because the victims have not been named. What the reports do establish is the shape of the threat, and the shape is new. The attackers are not only breaking into systems; they are poisoning the tools that developers use to build systems, and they are automating the work of finding and exploiting vulnerabilities faster than human defenders can respond. The next edition of Google's tracker, and the next FBI advisory, will show whether the defences are catching up, or whether the gap is still widening.

Source: Google Threat Intelligence Group via SiliconANGLE

Advertisement
Advertisement

Leave a Reply

Your email address will not be published. Required fields are marked *

Advertisement
Advertisement
About Us  |  Contact Us  |  Privacy Policy  |  Disclaimer  |  Terms and Conditions