Aisuru, the botnet responsible for a series of record-smashing distributed denial-of-service (DDoS) attacks this year, recently was overhauled to support a more low-key, lucrative and sustainable business: Renting hundreds of thousands of infected Internet of Things (IoT) devices to proxy services that help cybercriminals anonymize their traffic. Experts say a glut of proxies from Aisuru and other sources is fueling large-scale data harvesting efforts tied to various artificial intelligence (AI) projects, helping content scrapers evade detection by routing their traffic through residential connections that appear to be regular Internet users.

First identified in August 2024, Aisuru has spread to at least 700,000 IoT systems, such as poorly secured Internet routers and security cameras. Aisuru’s overlords have used their massive botnet to clobber targets with headline-grabbing DDoS attacks, flooding targeted hosts with blasts of junk requests from all infected systems simultaneously.

In June, Aisuru hit KrebsOnSecurity.com with a DDoS clocking at 6.3 terabits per second — the biggest attack that Google had ever mitigated at the time. In the weeks and months that followed, Aisuru’s operators demonstrated DDoS capabilities of nearly 30 terabits of data per second — well beyond the attack mitigation capabilities of most Internet destinations.

These digital sieges have been particularly disruptive this year for U.S.-based Internet service providers (ISPs), in part because Aisuru recently succeeded in taking over a large number of IoT devices in the United States. And when Aisuru launches attacks, the volume of outgoing traffic from infected systems on these ISPs is often so high that it can disrupt or degrade Internet service for adjacent (non-botted) customers of the ISPs.

“Multiple broadband access network operators have experienced significant operational impact due to outbound DDoS attacks in excess of 1.5Tb/sec launched from Aisuru botnet nodes residing on end-customer premises,” wrote Roland Dobbins, principal engineer at Netscout, in a recent executive summary on Aisuru. “Outbound/crossbound attack traffic exceeding 1Tb/sec from compromised customer premise equipment (CPE) devices has caused significant disruption to wireline and wireless broadband access networks. High-throughput attacks have caused chassis-based router line card failures.”

The incessant attacks from Aisuru have caught the attention of federal authorities in the United States and Europe (many of Aisuru’s victims are customers of ISPs and hosting providers based in Europe). Quite recently, some of the world’s largest ISPs have started informally sharing block lists identifying the rapidly shifting locations of the servers that the attackers use to control the activities of the botnet.

Experts say the Aisuru botmasters recently updated their malware so that compromised devices can more easily be rented to so-called “residential proxy” providers. These proxy services allow paying customers to route their Internet communications through someone else’s device, providing anonymity and the ability to appear as a regular Internet user in almost any major city worldwide.

From a website’s perspective, the IP traffic of a residential proxy network user appears to originate from the rented residential IP address, not from the proxy service customer. Proxy services can be used in a legitimate manner for several business purposes — such as price comparisons or sales intelligence. But they are massively abused for hiding cybercrime activity (think advertising fraud, credential stuffing) because they can make it difficult to trace malicious traffic to its original source.

And as we’ll see in a moment, this entire shadowy industry appears to be shifting its focus toward enabling aggressive content scraping activity that continuously feeds raw data into large language models (LLMs) built to support various AI projects.

‘INSANE’ GROWTH

Riley Kilmer is co-founder of spur.us, a service that tracks proxy networks. Kilmer said all of the top proxy services have grown substantially over the past six months.

“I just checked, and in the last 90 days we’ve seen 250 million unique residential proxy IPs,” Kilmer said. “That is insane. That is so high of a number, it’s unheard of. These proxies are absolutely everywhere now.”

Today, Spur says it is tracking an unprecedented spike in available proxies across all providers, including;

LUMINATI_PROXY    11,856,421
NETNUT_PROXY    10,982,458
ABCPROXY_PROXY    9,294,419
OXYLABS_PROXY     6,754,790
IPIDEA_PROXY     3,209,313
EARNFM_PROXY    2,659,913
NODEMAVEN_PROXY    2,627,851
INFATICA_PROXY    2,335,194
IPROYAL_PROXY    2,032,027
YILU_PROXY    1,549,155

Reached for comment about the apparent rapid growth in their proxy network, Oxylabs (#4 on Spur’s list) said while their proxy pool did grow recently, it did so at nowhere near the rate cited by Spur.

“We don’t systematically track other providers’ figures, and we’re not aware of any instances of 10× or 100× growth, especially when it comes to a few bigger companies that are legitimate businesses,” the company said in a written statement.

Bright Data was formerly known as Luminati Networks, the name that is currently at the top of Spur’s list of the biggest residential proxy networks. Bright Data likewise told KrebsOnSecurity that Spur’s current estimates of its proxy network are dramatically overstated and inaccurate.

“We did not actively initiate nor do we see any 10x or 100x expansion of our network, which leads me to believe that someone might be presenting these IPs as Bright Data’s in some way,” said Rony Shalit, Bright Data’s chief compliance and ethics officer. “In many cases in the past, due to us being the leading data collection proxy provider, IPs were falsely tagged as being part of our network, or while being used by other proxy providers for malicious activity.”

“Our network is only sourced from verified IP providers and a robust opt-in only residential peers, which we work hard and in complete transparency to obtain,” Shalit continued. “Every DC, ISP or SDK partner is reviewed and approved, and every residential peer must actively opt in to be part of our network.”

HK NETWORK

Even Spur acknowledges that Luminati and Oxylabs are unlike most other proxy services on their top proxy providers list, in that these providers actually adhere to “know-your-customer” policies, such as requiring video calls with all customers, and strictly blocking customers from reselling access.

Benjamin Brundage is founder of Synthient, a startup that helps companies detect proxy networks. Brundage said if there is increasing confusion around which proxy networks are the most worrisome, it’s because nearly all of these lesser-known proxy services have evolved into highly incestuous bandwidth resellers. What’s more, he said, some proxy providers do not appreciate being tracked and have been known to take aggressive steps to confuse systems that scan the Internet for residential proxy nodes.

Brundage said most proxy services today have created their own software development kit or SDK that other app developers can bundle with their code to earn revenue. These SDKs quietly modify the user’s device so that some portion of their bandwidth can be used to forward traffic from proxy service customers.

“Proxy providers have pools of constantly churning IP addresses,” he said. “These IP addresses are sourced through various means, such as bandwidth-sharing apps, botnets, Android SDKs, and more. These providers will often either directly approach resellers or offer a reseller program that allows users to resell bandwidth through their platform.”

Many SDK providers say they require full consent before allowing their software to be installed on end-user devices. Still, those opt-in agreements and consent checkboxes may be little more than a formality for cybercriminals like the Aisuru botmasters, who can earn a commission each time one of their infected devices is forced to install some SDK that enables one or more of these proxy services.

Depending on its structure, a single provider may operate hundreds of different proxy pools at a time — all maintained through other means, Brundage said.

“Often, you’ll see resellers maintaining their own proxy pool in addition to an upstream provider,” he said. “It allows them to market a proxy pool to high-value clients and offer an unlimited bandwidth plan for cheap reduce their own costs.”

Some proxy providers appear to be directly in league with botmasters. Brundage identified one proxy seller that was aggressively advertising cheap and plentiful bandwidth to content scraping companies. After scanning that provider’s pool of available proxies, Brundage said he found a one-to-one match with IP addresses he’d previously mapped to the Aisuru botnet.

Brundage says that by almost any measurement, the world’s largest residential proxy service is IPidea, a China-based proxy network. IPidea is #5 on Spur’s Top 10, and Brundage said its brands include ABCProxy (#3), Roxlabs, LunaProxy, PIA S5 Proxy, PyProxy, 922Proxy, 360Proxy, IP2World, and Cherry Proxy. Spur’s Kilmer said they also track Yilu Proxy (#10) as IPidea.

Brundage said all of these providers operate under a corporate umbrella known on the cybercrime forums as “HK Network.”

“The way it works is there’s this whole reseller ecosystem, where IPidea will be incredibly aggressive and approach all these proxy providers with the offer, ‘Hey, if you guys buy bandwidth from us, we’ll give you these amazing reseller prices,'” Brundage explained. “But they’re also very aggressive in recruiting resellers for their apps.”

Those apps include a range of low-cost and “free” virtual private networking (VPN) services that indeed allow users to enjoy a free VPN, but which also turn the user’s device into a traffic relay that can be rented to cybercriminals, or else parceled out to countless other proxy networks.

“They have all this bandwidth to offload,” Brundage said of IPidea and its sister networks. “And they can do it through their own platforms, or they go get resellers to do it for them by advertising on sketchy hacker forums to reach more people.”

One of IPidea’s core brands is 922S5Proxy, which is a not-so-subtle nod to the 911S5Proxy service that was hugely popular between 2015 and 2022. In July 2022, KrebsOnSecurity published a deep dive into 911S5Proxy’s origins and apparent owners in China. Less than a week later, 911S5Proxy announced it was closing down after the company’s servers were massively hacked.

That 2022 story named Yunhe Wang from Beijing as the apparent owner and/or manager of the 911S5 proxy service. In May 2024, the U.S. Department of Justice arrested Mr Wang, alleging that his network was used to steal billions of dollars from financial institutions, credit card issuers, and federal lending programs. At the same time, the U.S. Treasury Department announced sanctions against Wang and two other Chinese nationals for operating 911S5Proxy.

DATA SCRAPING FOR AI

In recent months, multiple experts who track botnet and proxy activity have shared that a great deal of content scraping which ultimately benefits AI companies is now leveraging these proxy networks to further obfuscate their aggressive data-slurping activity. That’s because by routing it through residential IP addresses, content scraping firms can make their traffic far trickier to filter out.

“It’s really difficult to block, because there’s a risk of blocking real people,” Spur’s Kilmer said of the LLM scraping activity that is fed through individual residential IP addresses, which are often shared by multiple customers at once.

Kilmer says the AI industry has brought a veneer of legitimacy to residential proxy business, which has heretofore mostly been associated with sketchy affiliate money making programs, automated abuse, and unwanted Internet traffic.

“Web crawling and scraping has always been a thing, but AI made it like a commodity, data that had to be collected,” Kilmer said. “Everybody wanted to monetize their own data pots, and how they monetize that is different across the board.”

Kilmer said many LLM-related scrapers rely on residential proxies in cases where the content provider has restricted access to their platform in some way, such as forcing interaction through an app, or keeping all content behind a login page with multi-factor authentication.

“Where the cost of data is out of reach — there is some exclusivity or reason they can’t access the data — they’ll turn to residential proxies so they look like a real person accessing that data,” Kilmer said of the content scraping efforts.

Aggressive AI crawlers increasingly are overloading community-maintained infrastructure, causing what amounts to persistent DDoS attacks on vital public resources. A report earlier this year from LibreNews found some open-source projects now see as much as 97 percent of their traffic originating from AI company bots, dramatically increasing bandwidth costs, service instability, and burdening already stretched-thin maintainers.

Cloudflare is now experimenting with tools that will allow content creators to charge a fee to AI crawlers to scrape their websites. The company’s “pay-per-crawl” feature is currently in a private beta, and it lets publishers set their own prices that bots must pay before scraping content.

On October 22, the social media and news network Reddit sued Oxylabs (PDF) and several other proxy providers, alleging that their systems enabled the mass-scraping of Reddit user content even though Reddit had taken steps to block such activity.

“Recognizing that Reddit denies scrapers like them access to its site, Defendants scrape the data from Google’s search results instead,” the lawsuit alleges. “They do so by masking their identities, hiding their locations, and disguising their web scrapers as regular people (among other techniques) to circumvent or bypass the security restrictions meant to stop them.”

Denas Grybauskas, chief governance and strategy officer at Oxylabs, said the company was shocked and disappointed by the lawsuit.

“Reddit has made no attempt to speak with us directly or communicate any potential concerns,” Grybauskas said in a written statement. “Oxylabs has always been and will continue to be a pioneer and an industry leader in public data collection, and it will not hesitate to defend itself against these allegations. Oxylabs’ position is that no company should claim ownership of public data that does not belong to them. It is possible that it is just an attempt to sell the same public data at an inflated price.”

As big and powerful as Aisuru may be, it is hardly the only botnet that is contributing to the overall broad availability of residential proxies. For example, on June 5 the FBI’s Internet Crime Complaint Center warned that an IoT malware threat dubbed BADBOX 2.0 had compromised millions of smart-TV boxes, digital projectors, vehicle infotainment units, picture frames, and other IoT devices.

In July, Google filed a lawsuit in New York federal court against the Badbox botnet’s alleged perpetrators. Google said the Badbox 2.0 botnet “compromised more than 10 million uncertified devices running Android’s open-source software, which lacks Google’s security protections. Cybercriminals infected these devices with pre-installed malware and exploited them to conduct large-scale ad fraud and other digital crimes.”

A FAMILIAR DOMAIN NAME

Brundage said the Aisuru botmasters have their own SDK, and for some reason part of its code tells many newly-infected systems to query the domain name fuckbriankrebs[.]com. This may be little more than an elaborate “screw you” to this site’s author: One of the botnet’s alleged partners goes by the handle “Forky,” and was identified in June by KrebsOnSecurity as a young man from Sao Paulo, Brazil.

Brundage noted that only systems infected with Aisuru’s Android SDK will be forced to resolve the domain. Initially, there was some discussion about whether the domain might have some utility as a “kill switch” capable of disrupting the botnet’s operations, although Brundage and others interviewed for this story say that is unlikely.

For one thing, they said, if the domain was somehow critical to the operation of the botnet, why was it still unregistered and actively for-sale? Why indeed, we asked. Happily, the domain name was deftly snatched up last week by Philippe Caturegli, “chief hacking officer” for the security intelligence company Seralys.

Caturegli enabled a passive DNS server on that domain and within a few hours received more than 700,000 requests for unique subdomains on fuckbriankrebs[.]com.

But even with that visibility into Aisuru, it is difficult to use this domain check-in feature to measure its true size, Brundage said. After all, he said, the systems that are phoning home to the domain are only a small portion of the overall botnet.

“The bots are hardcoded to just spam lookups on the subdomains,” he said. “So anytime an infection occurs or it runs in the background, it will do one of those DNS queries.”

The domain fuckbriankrebs[.]com has a storied history. On its initial launch in 2009, it was used to spread malicious software by the Cutwail spam botnet. In 2011, the domain was involved in a notable DDoS against this website from a botnet powered by Russkill (a.k.a. “Dirt Jumper”).

Domaintools.com finds that in 2015, fuckbriankrebs[.]com was registered to an email address attributed to David “Abdilo” Crees, a 27-year-old Australian man sentenced in May 2025 to time served for cybercrime convictions related to the Lizard Squad hacking group.

Update, Nov. 1, 2025, 10:25 a.m. ET: An earlier version of this story erroneously cited Spur’s proxy numbers from earlier this year; Spur said those numbers conflated residential proxies — which are rotating and attached to real end-user devices — with “ISP proxies” located at AT&T. ISP proxies, Spur said, involve tricking an ISP into routing a large number of IP addresses that are resold as far more static datacenter proxies.

At least 18 popular JavaScript code packages that are collectively downloaded more than two billion times each week were briefly compromised with malicious software today, after a developer involved in maintaining the projects was phished. The attack appears to have been quickly contained and was narrowly focused on stealing cryptocurrency. But experts warn that a similar attack with a slightly more nefarious payload could lead to a disruptive malware outbreak that is far more difficult to detect and restrain.

Aikido is a security firm in Belgium that monitors new code updates to major open-source code repositories, scanning any code updates for suspicious and malicious code. In a blog post published today, Aikido said its systems found malicious code had been added to at least 18 widely-used code libraries available on NPM (short for) “Node Package Manager,” which acts as a central hub for JavaScript development and the latest updates to widely-used JavaScript components.

JavaScript is a powerful web-based scripting language used by countless websites to build a more interactive experience with users, such as entering data into a form. But there’s no need for each website developer to build a program from scratch for entering data into a form when they can just reuse already existing packages of code at NPM that are specifically designed for that purpose.

Unfortunately, if cybercriminals manage to phish NPM credentials from developers, they can introduce malicious code that allows attackers to fundamentally control what people see in their web browser when they visit a website that uses one of the affected code libraries.

According to Aikido, the attackers injected a piece of code that silently intercepts cryptocurrency activity in the browser, “manipulates wallet interactions, and rewrites payment destinations so that funds and approvals are redirected to attacker-controlled accounts without any obvious signs to the user.”

“This malware is essentially a browser-based interceptor that hijacks both network traffic and application APIs,” Aikido researcher Charlie Eriksen wrote. “What makes it dangerous is that it operates at multiple layers: Altering content shown on websites, tampering with API calls, and manipulating what users’ apps believe they are signing. Even if the interface looks correct, the underlying transaction can be redirected in the background.”

Aikido said it used the social network Bsky to notify the affected developer, Josh Junon, who quickly replied that he was aware of having just been phished. The phishing email that Junon fell for was part of a larger campaign that spoofed NPM and told recipients they were required to update their two-factor authentication (2FA) credentials. The phishing site mimicked NPM’s login page, and intercepted Junon’s credentials and 2FA token. Once logged in, the phishers then changed the email address on file for Junon’s NPM account, temporarily locking him out.

Junon also issued a mea culpa on HackerNews, telling the community’s coder-heavy readership, “Hi, yep I got pwned.”

“It looks and feels a bit like a targeted attack,” Junon wrote. “Sorry everyone, very embarrassing.”

Philippe Caturegli, “chief hacking officer” at the security consultancy Seralys, observed that the attackers appear to have registered their spoofed website — npmjs[.]help — just two days before sending the phishing email. The spoofed website used services from dnsexit[.]com, a “dynamic DNS” company that also offers “100% free” domain names that can instantly be pointed at any IP address controlled by the user.

Caturegli said it’s remarkable that the attackers in this case were not more ambitious or malicious with their code modifications.

“The crazy part is they compromised billions of websites and apps just to target a couple of cryptocurrency things,” he said. “This was a supply chain attack, and it could easily have been something much worse than crypto harvesting.”

Aikido’s Eriksen agreed, saying countless websites dodged a bullet because this incident was handled in a matter of hours. As an example of how these supply-chain attacks can escalate quickly, Eriksen pointed to another compromise of an NPM developer in late August that added malware to “nx,” an open-source code development toolkit with as many as six million weekly downloads.

In the nx compromise, the attackers introduced code that scoured the user’s device for authentication tokens from programmer destinations like GitHub and NPM, as well as SSH and API keys. But instead of sending those stolen credentials to a central server controlled by the attackers, the malicious code created a new public repository in the victim’s GitHub account, and published the stolen data there for all the world to see and download.

Eriksen said coding platforms like GitHub and NPM should be doing more to ensure that any new code commits for broadly-used packages require a higher level of attestation that confirms the code in question was in fact submitted by the person who owns the account, and not just by that person’s account.

“More popular packages should require attestation that it came through trusted provenance and not just randomly from some location on the Internet,” Eriksen said. “Where does the package get uploaded from, by GitHub in response to a new pull request into the main branch, or somewhere else? In this case, they didn’t compromise the target’s GitHub account. They didn’t touch that. They just uploaded a modified version that didn’t come where it’s expected to come from.”

Eriksen said code repository compromises can be devastating for developers, many of whom end up abandoning their projects entirely after such an incident.

“It’s unfortunate because one thing we’ve seen is people have their projects get compromised and they say, ‘You know what, I don’t have the energy for this and I’m just going to deprecate the whole package,'” Eriksen said.

Kevin Beaumont, a frequently quoted security expert who writes about security incidents at the blog doublepulsar.com, has been following this story closely today in frequent updates to his account on Mastodon. Beaumont said the incident is a reminder that much of the planet still depends on code that is ultimately maintained by an exceedingly small number of people who are mostly overburdened and under-resourced.

“For about the past 15 years every business has been developing apps by pulling in 178 interconnected libraries written by 24 people in a shed in Skegness,” Beaumont wrote on Mastodon. “For about the past 2 years orgs have been buying AI vibe coding tools, where some exec screams ‘make online shop’ into a computer and 389 libraries are added and an app is farted out. The output = if you want to own the world’s companies, just phish one guy in Skegness.”

Aikido recently launched a product that aims to help development teams ensure that every code library used is checked for malware before it can be used or installed. Nicholas Weaver, a researcher with the International Computer Science Institute, a nonprofit in Berkeley, Calif., said Aikido’s new offering exists because many organizations are still one successful phishing attack away from a supply-chain nightmare.

Weaver said these types of supply-chain compromises will continue as long as people responsible for maintaining widely-used code continue to rely on phishable forms of 2FA.

“NPM should only support phish-proof authentication,” Weaver said, referring to physical security keys that are phish-proof — meaning that even if phishers manage to steal your username and password, they still can’t log in to your account without also possessing that physical key.

“All critical infrastructure needs to use phish-proof 2FA, and given the dependencies in modern software, archives such as NPM are absolutely critical infrastructure,” Weaver said. “That NPM does not require that all contributor accounts use security keys or similar 2FA methods should be considered negligence.”

Marko Elez, a 25-year-old employee at Elon Musk’s Department of Government Efficiency (DOGE), has been granted access to sensitive databases at the U.S. Social Security Administration, the Treasury and Justice departments, and the Department of Homeland Security. So it should fill all Americans with a deep sense of confidence to learn that Mr. Elez over the weekend inadvertently published a private key that allowed anyone to interact directly with more than four dozen large language models (LLMs) developed by Musk’s artificial intelligence company xAI.

On July 13, Mr. Elez committed a code script to GitHub called “agent.py” that included a private application programming interface (API) key for xAI. The inclusion of the private key was first flagged by GitGuardian, a company that specializes in detecting and remediating exposed secrets in public and proprietary environments. GitGuardian’s systems constantly scan GitHub and other code repositories for exposed API keys, and fire off automated alerts to affected users.

Philippe Caturegli, “chief hacking officer” at the security consultancy Seralys, said the exposed API key allowed access to at least 52 different LLMs used by xAI. The most recent LLM in the list was called “grok-4-0709” and was created on July 9, 2025.

Grok, the generative AI chatbot developed by xAI and integrated into Twitter/X, relies on these and other LLMs (a query to Grok before publication shows Grok currently uses Grok-3, which was launched in Feburary 2025). Earlier today, xAI announced that the Department of Defense will begin using Grok as part of a contract worth up to $200 million. The contract award came less than a week after Grok began spewing antisemitic rants and invoking Adolf Hitler.

Mr. Elez did not respond to a request for comment. The code repository containing the private xAI key was removed shortly after Caturegli notified Elez via email. However, Caturegli said the exposed API key still works and has not yet been revoked.

“If a developer can’t keep an API key private, it raises questions about how they’re handling far more sensitive government information behind closed doors,” Caturegli told KrebsOnSecurity.

Prior to joining DOGE, Marko Elez worked for a number of Musk’s companies. His DOGE career began at the Department of the Treasury, and a legal battle over DOGE’s access to Treasury databases showed Elez was sending unencrypted personal information in violation of the agency’s policies.

While still at Treasury, Elez resigned after The Wall Street Journal linked him to social media posts that advocated racism and eugenics. When Vice President J.D. Vance lobbied for Elez to be rehired, President Trump agreed and Musk reinstated him.

Since his re-hiring as a DOGE employee, Elez has been granted access to databases at one federal agency after another. TechCrunch reported in February 2025 that he was working at the Social Security Administration. In March, Business Insider found Elez was part of a DOGE detachment assigned to the Department of Labor.

In April, The New York Times reported that Elez held positions at the U.S. Customs and Border Protection and the Immigration and Customs Enforcement (ICE) bureaus, as well as the Department of Homeland Security. The Washington Post later reported that Elez, while serving as a DOGE advisor at the Department of Justice, had gained access to the Executive Office for Immigration Review’s Courts and Appeals System (EACS).

Elez is not the first DOGE worker to publish internal API keys for xAI: In May, KrebsOnSecurity detailed how another DOGE employee leaked a private xAI key on GitHub for two months, exposing LLMs that were custom made for working with internal data from Musk’s companies, including SpaceX, Tesla and Twitter/X.

Caturegli said it’s difficult to trust someone with access to confidential government systems when they can’t even manage the basics of operational security.

“One leak is a mistake,” he said. “But when the same type of sensitive key gets exposed again and again, it’s not just bad luck, it’s a sign of deeper negligence and a broken security culture.”

An employee at Elon Musk’s artificial intelligence company xAI leaked a private key on GitHub that for the past two months could have allowed anyone to query private xAI large language models (LLMs) which appear to have been custom made for working with internal data from Musk’s companies, including SpaceX, Tesla and Twitter/X, KrebsOnSecurity has learned.

Philippe Caturegli, “chief hacking officer” at the security consultancy Seralys, was the first to publicize the leak of credentials for an x.ai application programming interface (API) exposed in the GitHub code repository of a technical staff member at xAI.

Caturegli’s post on LinkedIn caught the attention of researchers at GitGuardian, a company that specializes in detecting and remediating exposed secrets in public and proprietary environments. GitGuardian’s systems constantly scan GitHub and other code repositories for exposed API keys, and fire off automated alerts to affected users.

GitGuardian’s Eric Fourrier told KrebsOnSecurity the exposed API key had access to several unreleased models of Grok, the AI chatbot developed by xAI. In total, GitGuardian found the key had access to at least 60 fine-tuned and private LLMs.

“The credentials can be used to access the X.ai API with the identity of the user,” GitGuardian wrote in an email explaining their findings to xAI. “The associated account not only has access to public Grok models (grok-2-1212, etc) but also to what appears to be unreleased (grok-2.5V), development (research-grok-2p5v-1018), and private models (tweet-rejector, grok-spacex-2024-11-04).”

Fourrier found GitGuardian had alerted the xAI employee about the exposed API key nearly two months ago — on March 2. But as of April 30, when GitGuardian directly alerted xAI’s security team to the exposure, the key was still valid and usable. xAI told GitGuardian to report the matter through its bug bounty program at HackerOne, but just a few hours later the repository containing the API key was removed from GitHub.

“It looks like some of these internal LLMs were fine-tuned on SpaceX data, and some were fine-tuned with Tesla data,” Fourrier said. “I definitely don’t think a Grok model that’s fine-tuned on SpaceX data is intended to be exposed publicly.”

xAI did not respond to a request for comment. Nor did the 28-year-old xAI technical staff member whose key was exposed.

Carole Winqwist, chief marketing officer at GitGuardian, said giving potentially hostile users free access to private LLMs is a recipe for disaster.

“If you’re an attacker and you have direct access to the model and the back end interface for things like Grok, it’s definitely something you can use for further attacking,” she said. “An attacker could it use for prompt injection, to tweak the (LLM) model to serve their purposes, or try to implant code into the supply chain.”

The inadvertent exposure of internal LLMs for xAI comes as Musk’s so-called Department of Government Efficiency (DOGE) has been feeding sensitive government records into artificial intelligence tools. In February, The Washington Post reported DOGE officials were feeding data from across the Education Department into AI tools to probe the agency’s programs and spending.

The Post said DOGE plans to replicate this process across many departments and agencies, accessing the back-end software at different parts of the government and then using AI technology to extract and sift through information about spending on employees and programs.

“Feeding sensitive data into AI software puts it into the possession of a system’s operator, increasing the chances it will be leaked or swept up in cyberattacks,” Post reporters wrote.

Wired reported in March that DOGE has deployed a proprietary chatbot called GSAi to 1,500 federal workers at the General Services Administration, part of an effort to automate tasks previously done by humans as DOGE continues its purge of the federal workforce.

A Reuters report last month said Trump administration officials told some U.S. government employees that DOGE is using AI to surveil at least one federal agency’s communications for hostility to President Trump and his agenda. Reuters wrote that the DOGE team has heavily deployed Musk’s Grok AI chatbot as part of their work slashing the federal government, although Reuters said it could not establish exactly how Grok was being used.

Caturegli said while there is no indication that federal government or user data could be accessed through the exposed x.ai API key, these private models are likely trained on proprietary data and may unintentionally expose details related to internal development efforts at xAI, Twitter, or SpaceX.

“The fact that this key was publicly exposed for two months and granted access to internal models is concerning,” Caturegli said. “This kind of long-lived credential exposure highlights weak key management and insufficient internal monitoring, raising questions about safeguards around developer access and broader operational security.”