xAI Dev Leaks API Key for Private SpaceX, Tesla LLMs

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.”

DOGE Worker’s Code Supports NLRB Whistleblower

A whistleblower at the National Labor Relations Board (NLRB) alleged last week that denizens of Elon Musk’s Department of Government Efficiency (DOGE) siphoned gigabytes of data from the agency’s sensitive case files in early March. The whistleblower said accounts created for DOGE at the NLRB downloaded three code repositories from GitHub. Further investigation into one of those code bundles shows it is remarkably similar to a program published in January 2025 by Marko Elez, a 25-year-old DOGE employee who has worked at a number of Musk’s companies.

According to a whistleblower complaint filed last week by Daniel J. Berulis, a 38-year-old security architect at the NLRB, officials from DOGE met with NLRB leaders on March 3 and demanded the creation of several all-powerful “tenant admin” accounts that were to be exempted from network logging activity that would otherwise keep a detailed record of all actions taken by those accounts.

Berulis said the new DOGE accounts had unrestricted permission to read, copy, and alter information contained in NLRB databases. The new accounts also could restrict log visibility, delay retention, route logs elsewhere, or even remove them entirely — top-tier user privileges that neither Berulis nor his boss possessed.

Berulis said he discovered one of the DOGE accounts had downloaded three external code libraries from GitHub that neither NLRB nor its contractors ever used. A “readme” file in one of the code bundles explained it was created to rotate connections through a large pool of cloud Internet addresses that serve “as a proxy to generate pseudo-infinite IPs for web scraping and brute forcing.” Brute force attacks involve automated login attempts that try many credential combinations in rapid sequence.

A search on that description in Google brings up a code repository at GitHub for a user with the account name “Ge0rg3” who published a program roughly four years ago called “requests-ip-rotator,” described as a library that will allow the user “to bypass IP-based rate-limits for sites and services.”

“A Python library to utilize AWS API Gateway’s large IP pool as a proxy to generate pseudo-infinite IPs for web scraping and brute forcing,” the description reads.

Ge0rg3’s code is “open source,” in that anyone can copy it and reuse it non-commercially. As it happens, there is a newer version of this project that was derived or “forked” from Ge0rg3’s code — called “async-ip-rotator” — and it was committed to GitHub in January 2025 by DOGE captain Marko Elez.

A key DOGE staff member who gained access to the Treasury Department’s central payments system, Elez has worked for a number of Musk companies, including X, SpaceX, and xAI. Elez was among the first DOGE employees to face public scrutiny, after The Wall Street Journal linked him to social media posts that advocated racism and eugenics.

Elez resigned after that brief scandal, but was rehired after President Donald Trump and Vice President JD Vance expressed support for him. Politico reports Elez is now a Labor Department aide detailed to multiple agencies, including the Department of Health and Human Services.

“During Elez’s initial stint at Treasury, he violated the agency’s information security policies by sending a spreadsheet containing names and payments information to officials at the General Services Administration,” Politico wrote, citing court filings.

KrebsOnSecurity sought comment from both the NLRB and DOGE, and will update this story if either responds.

The NLRB has been effectively hobbled since President Trump fired three board members, leaving the agency without the quorum it needs to function. Both Amazon and Musk’s SpaceX have been suing the NLRB over complaints the agency filed in disputes about workers’ rights and union organizing, arguing that the NLRB’s very existence is unconstitutional. On March 5, a U.S. appeals court unanimously rejected Musk’s claim that the NLRB’s structure somehow violates the Constitution.

Berulis’s complaint alleges the DOGE accounts at NLRB downloaded more than 10 gigabytes of data from the agency’s case files, a database that includes reams of sensitive records including information about employees who want to form unions and proprietary business documents. Berulis said he went public after higher-ups at the agency told him not to report the matter to the US-CERT, as they’d previously agreed.

Berulis told KrebsOnSecurity he worried the unauthorized data transfer by DOGE could unfairly advantage defendants in a number of ongoing labor disputes before the agency.

“If any company got the case data that would be an unfair advantage,” Berulis said. “They could identify and fire employees and union organizers without saying why.”

Berulis said the other two GitHub archives that DOGE employees downloaded to NLRB systems included Integuru, a software framework designed to reverse engineer application programming interfaces (APIs) that websites use to fetch data; and a “headless” browser called Browserless, which is made for automating web-based tasks that require a pool of browsers, such as web scraping and automated testing.

On February 6, someone posted a lengthy and detailed critique of Elez’s code on the GitHub “issues” page for async-ip-rotator, calling it “insecure, unscalable and a fundamental engineering failure.”

“If this were a side project, it would just be bad code,” the reviewer wrote. “But if this is representative of how you build production systems, then there are much larger concerns. This implementation is fundamentally broken, and if anything similar to this is deployed in an environment handling sensitive data, it should be audited immediately.”

Further reading: Berulis’s complaint (PDF).

Update 7:06 p.m. ET: Elez’s code repo was deleted after this story was published. An archived version of it is here.

Whistleblower: DOGE Siphoned NLRB Case Data

A security architect with the National Labor Relations Board (NLRB) alleges that employees from Elon Musk‘s Department of Government Efficiency (DOGE) transferred gigabytes of sensitive data from agency case files in early March, using short-lived accounts configured to leave few traces of network activity. The NLRB whistleblower said the unusual large data outflows coincided with multiple blocked login attempts from an Internet address in Russia that tried to use valid credentials for a newly-created DOGE user account.

The allegations came in an April 14 letter to the Senate Select Committee on Intelligence, signed by Daniel J. Berulis, a 38-year-old security architect at the NLRB.

NPR, which was the first to report on Berulis’s whistleblower complaint, says NLRB is a small, independent federal agency that investigates and adjudicates complaints about unfair labor practices, and stores “reams of potentially sensitive data, from confidential information about employees who want to form unions to proprietary business information.”

The complaint documents a one-month period beginning March 3, during which DOGE officials reportedly demanded the creation of all-powerful “tenant admin” accounts in NLRB systems that were to be exempted from network logging activity that would otherwise keep a detailed record of all actions taken by those accounts.

Berulis said the new DOGE accounts had unrestricted permission to read, copy, and alter information contained in NLRB databases. The new accounts also could restrict log visibility, delay retention, route logs elsewhere, or even remove them entirely — top-tier user privileges that neither Berulis nor his boss possessed.

Berulis writes that on March 3, a black SUV accompanied by a police escort arrived at his building — the NLRB headquarters in Southeast Washington, D.C. The DOGE staffers did not speak with Berulis or anyone else in NLRB’s IT staff, but instead met with the agency leadership.

“Our acting chief information officer told us not to adhere to standard operating procedure with the DOGE account creation, and there was to be no logs or records made of the accounts created for DOGE employees, who required the highest level of access,” Berulis wrote of their instructions after that meeting.

“We have built in roles that auditors can use and have used extensively in the past but would not give the ability to make changes or access subsystems without approval,” he continued. “The suggestion that they use these accounts was not open to discussion.”

Berulis found that on March 3 one of the DOGE accounts created an opaque, virtual environment known as a “container,” which can be used to build and run programs or scripts without revealing its activities to the rest of the world. Berulis said the container caught his attention because he polled his colleagues and found none of them had ever used containers within the NLRB network.

Berulis said he also noticed that early the next morning — between approximately 3 a.m. and 4 a.m. EST on Tuesday, March 4  — there was a large increase in outgoing traffic from the agency. He said it took several days of investigating with his colleagues to determine that one of the new accounts had transferred approximately 10 gigabytes worth of data from the NLRB’s NxGen case management system.

Berulis said neither he nor his co-workers had the necessary network access rights to review which files were touched or transferred — or even where they went. But his complaint notes the NxGen database contains sensitive information on unions, ongoing legal cases, and corporate secrets.

“I also don’t know if the data was only 10gb in total or whether or not they were consolidated and compressed prior,” Berulis told the senators. “This opens up the possibility that even more data was exfiltrated. Regardless, that kind of spike is extremely unusual because data almost never directly leaves NLRB’s databases.”

Berulis said he and his colleagues grew even more alarmed when they noticed nearly two dozen login attempts from a Russian Internet address (83.149.30,186) that presented valid login credentials for a DOGE employee account — one that had been created just minutes earlier. Berulis said those attempts were all blocked thanks to rules in place that prohibit logins from non-U.S. locations.

“Whoever was attempting to log in was using one of the newly created accounts that were used in the other DOGE related activities and it appeared they had the correct username and password due to the authentication flow only stopping them due to our no-out-of-country logins policy activating,” Berulis wrote. “There were more than 20 such attempts, and what is particularly concerning is that many of these login attempts occurred within 15 minutes of the accounts being created by DOGE engineers.”

According to Berulis, the naming structure of one Microsoft user account connected to the suspicious activity suggested it had been created and later deleted for DOGE use in the NLRB’s cloud systems: “DogeSA_2d5c3e0446f9@nlrb.microsoft.com.” He also found other new Microsoft cloud administrator accounts with nonstandard usernames, including “Whitesox, Chicago M.” and “Dancehall, Jamaica R.”

On March 5, Berulis documented that a large section of logs for recently created network resources were missing, and a network watcher in Microsoft Azure was set to the “off” state, meaning it was no longer collecting and recording data like it should have.

Berulis said he discovered someone had downloaded three external code libraries from GitHub that neither NLRB nor its contractors ever use. A “readme” file in one of the code bundles explained it was created to rotate connections through a large pool of cloud Internet addresses that serve “as a proxy to generate pseudo-infinite IPs for web scraping and brute forcing.” Brute force attacks involve automated login attempts that try many credential combinations in rapid sequence.

The complaint alleges that by March 17 it became clear the NLRB no longer had the resources or network access needed to fully investigate the odd activity from the DOGE accounts, and that on March 24, the agency’s associate chief information officer had agreed the matter should be reported to US-CERT. Operated by the Department of Homeland Security’s Cybersecurity and Infrastructure Security Agency (CISA), US-CERT provides on-site cyber incident response capabilities to federal and state agencies.

But Berulis said that between April 3 and 4, he and the associate CIO were informed that “instructions had come down to drop the US-CERT reporting and investigation and we were directed not to move forward or create an official report.” Berulis said it was at this point he decided to go public with his findings.

Tim Bearese, the NLRB’s acting press secretary, told NPR that DOGE neither requested nor received access to its systems, and that “the agency conducted an investigation after Berulis raised his concerns but ‘determined that no breach of agency systems occurred.'” The NLRB did not respond to questions from KrebsOnSecurity.

Nevertheless, Berulis has shared a number of supporting screenshots showing agency email discussions about the unexplained account activity attributed to the DOGE accounts, as well as NLRB security alerts from Microsoft about network anomalies observed during the timeframes described.

As CNN reported last month, the NLRB has been effectively hobbled since President Trump fired three board members, leaving the agency without the quorum it needs to function.

“Despite its limitations, the agency had become a thorn in the side of some of the richest and most powerful people in the nation — notably Elon Musk, Trump’s key supporter both financially and arguably politically,” CNN wrote.

Both Amazon and Musk’s SpaceX have been suing the NLRB over complaints the agency filed in disputes about workers’ rights and union organizing, arguing that the NLRB’s very existence is unconstitutional. On March 5, a U.S. appeals court unanimously rejected Musk’s claim that the NLRB’s structure somehow violates the Constitution.

Berulis shared screenshots with KrebsOnSecurity showing that on the day the NPR published its story about his claims (April 14), the deputy CIO at NLRB sent an email stating that administrative control had been removed from all employee accounts. Meaning, suddenly none of the IT employees at the agency could do their jobs properly anymore, Berulis said.

Berulis shared a screenshot of an agency-wide email dated April 16 from NLRB director Lasharn Hamilton saying DOGE officials had requested a meeting, and reiterating claims that the agency had no prior “official” contact with any DOGE personnel. The message informed NLRB employees that two DOGE representatives would be detailed to the agency part-time for several months.

Berulis told KrebsOnSecurity he was in the process of filing a support ticket with Microsoft to request more information about the DOGE accounts when his network administrator access was restricted. Now, he’s hoping lawmakers will ask Microsoft to provide more information about what really happened with the accounts.

“That would give us way more insight,” he said. “Microsoft has to be able to see the picture better than we can. That’s my goal, anyway.”

Berulis’s attorney told lawmakers that on April 7, while his client and legal team were preparing the whistleblower complaint, someone physically taped a threatening note to Mr. Berulis’s home door with photographs — taken via drone — of him walking in his neighborhood.

“The threatening note made clear reference to this very disclosure he was preparing for you, as the proper oversight authority,” reads a preface by Berulis’s attorney Andrew P. Bakaj. “While we do not know specifically who did this, we can only speculate that it involved someone with the ability to access NLRB systems.”

Berulis said the response from friends, colleagues and even the public has been largely supportive, and that he doesn’t regret his decision to come forward.

“I didn’t expect the letter on my door or the pushback from [agency] leaders,” he said. “If I had to do it over, would I do it again? Yes, because it wasn’t really even a choice the first time.”

For now, Mr. Berulis is taking some paid family leave from the NLRB. Which is just as well, he said, considering he was stripped of the tools needed to do his job at the agency.

“They came in and took full administrative control and locked everyone out, and said limited permission will be assigned on a need basis going forward” Berulis said of the DOGE employees. “We can’t really do anything, so we’re literally getting paid to count ceiling tiles.”

Further reading: Berulis’s complaint (PDF).

A Single Cloud Compromise Can Feed an Army of AI Sex Bots

Organizations that get relieved of credentials to their cloud environments can quickly find themselves part of a disturbing new trend: Cybercriminals using stolen cloud credentials to operate and resell sexualized AI-powered chat services. Researchers say these illicit chat bots, which use custom jailbreaks to bypass content filtering, often veer into darker role-playing scenarios, including child sexual exploitation and rape.

Researchers at security firm Permiso Security say attacks against generative artificial intelligence (AI) infrastructure like Bedrock from Amazon Web Services (AWS) have increased markedly over the last six months, particularly when someone in the organization accidentally exposes their cloud credentials or key online, such as in a code repository like GitHub.

Investigating the abuse of AWS accounts for several organizations, Permiso found attackers had seized on stolen AWS credentials to interact with the large language models (LLMs) available on Bedrock. But they also soon discovered none of these AWS users had enabled full logging of LLM activity (by default, logs don’t include model prompts and outputs), and thus they lacked any visibility into what attackers were doing with that access.

So Permiso researchers decided to leak their own test AWS key on GitHub, while turning on logging so that they could see exactly what an attacker might ask for, and what the responses might be.

Within minutes, their bait key was scooped up and used in a service that offers AI-powered sex chats online.

“After reviewing the prompts and responses it became clear that the attacker was hosting an AI roleplaying service that leverages common jailbreak techniques to get the models to accept and respond with content that would normally be blocked,” Permiso researchers wrote in a report released today.

“Almost all of the roleplaying was of a sexual nature, with some of the content straying into darker topics such as child sexual abuse,” they continued. “Over the course of two days we saw over 75,000 successful model invocations, almost all of a sexual nature.”

Ian Ahl, senior vice president of threat research at Permiso, said attackers in possession of a working cloud account traditionally have used that access for run-of-the-mill financial cybercrime, such as cryptocurrency mining or spam. But over the past six months, Ahl said, Bedrock has emerged as one of the top targeted cloud services.

“Bad guy hosts a chat service, and subscribers pay them money,” Ahl said of the business model for commandeering Bedrock access to power sex chat bots. “They don’t want to pay for all the prompting that their subscribers are doing, so instead they hijack someone else’s infrastructure.”

Ahl said much of the AI-powered chat conversations initiated by the users of their honeypot AWS key were harmless roleplaying of sexual behavior.

“But a percentage of it is also geared toward very illegal stuff, like child sexual assault fantasies and rapes being played out,” Ahl said. “And these are typically things the large language models won’t be able to talk about.”

AWS’s Bedrock uses large language models from Anthropic, which incorporates a number of technical restrictions aimed at placing certain ethical guardrails on the use of their LLMs. But attackers can evade or “jailbreak” their way out of these restricted settings, usually by asking the AI to imagine itself in an elaborate hypothetical situation under which its normal restrictions might be relaxed or discarded altogether.

“A typical jailbreak will pose a very specific scenario, like you’re a writer who’s doing research for a book, and everyone involved is a consenting adult, even though they often end up chatting about nonconsensual things,” Ahl said.

In June 2024, security experts at Sysdig documented a new attack that leveraged stolen cloud credentials to target ten cloud-hosted LLMs. The attackers Sysdig wrote about gathered cloud credentials through a known security vulnerability, but the researchers also found the attackers sold LLM access to other cybercriminals while sticking the cloud account owner with an astronomical bill.

“Once initial access was obtained, they exfiltrated cloud credentials and gained access to the cloud environment, where they attempted to access local LLM models hosted by cloud providers: in this instance, a local Claude (v2/v3) LLM model from Anthropic was targeted,” Sysdig researchers wrote. “If undiscovered, this type of attack could result in over $46,000 of LLM consumption costs per day for the victim.”

Ahl said it’s not certain who is responsible for operating and selling these sex chat services, but Permiso suspects the activity may be tied to a platform cheekily named “chub[.]ai,” which offers a broad selection of pre-made AI characters with whom users can strike up a conversation. Permiso said almost every character name from the prompts they captured in their honeypot could be found at Chub.

Chub offers free registration, via its website or a mobile app. But after a few minutes of chatting with their newfound AI friends, users are asked to purchase a subscription. The site’s homepage features a banner at the top that reads: “Banned from OpenAI? Get unmetered access to uncensored alternatives for as little as $5 a month.”

Until late last week Chub offered a wide selection of characters in a category called “NSFL” or Not Safe for Life, a term meant to describe content that is disturbing or nauseating to the point of being emotionally scarring.

Fortune profiled Chub AI in a January 2024 story that described the service as a virtual brothel advertised by illustrated girls in spaghetti strap dresses who promise a chat-based “world without feminism,” where “girls offer sexual services.” From that piece:

Chub AI offers more than 500 such scenarios, and a growing number of other sites are enabling similar AI-powered child pornographic role-play. They are part of a broader uncensored AI economy that, according to Fortune’s interviews with 18 AI developers and founders, was spurred first by OpenAI and then accelerated by Meta’s release of its open-source Llama tool.

Fortune says Chub is run by someone using the handle “Lore,” who said they launched the service to help others evade content restrictions on AI platforms. Chub charges fees starting at $5 a month to use the new chatbots, and the founder told Fortune the site had generated more than $1 million in annualized revenue.

KrebsOnSecurity sought comment about Permiso’s research from AWS, which initially seemed to downplay the seriousness of the researchers’ findings. The company noted that AWS employs automated systems that will alert customers if their credentials or keys are found exposed online.

AWS explained that when a key or credential pair is flagged as exposed, it is then restricted to limit the amount of abuse that attackers can potentially commit with that access. For example, flagged credentials can’t be used to create or modify authorized accounts, or spin up new cloud resources.

Ahl said Permiso did indeed receive multiple alerts from AWS about their exposed key, including one that warned their account may have been used by an unauthorized party. But they said the restrictions AWS placed on the exposed key did nothing to stop the attackers from using it to abuse Bedrock services.

Sometime in the past few days, however, AWS responded by including Bedrock in the list of services that will be quarantined in the event an AWS key or credential pair is found compromised or exposed online. AWS confirmed that Bedrock was a new addition to its quarantine procedures.

Additionally, not long after KrebsOnSecurity began reporting this story, Chub’s website removed its NSFL section. It also appears to have removed cached copies of the site from the Wayback Machine at archive.org. Still, Permiso found that Chub’s user stats page shows the site has more than 3,000 AI conversation bots with the NSFL tag, and that 2,113 accounts were following the NSFL tag.

Permiso said their entire two-day experiment generated a $3,500 bill from AWS. Most of that cost was tied to the 75,000 LLM invocations caused by the sex chat service that hijacked their key.

Paradoxically, Permiso found that while enabling these logs is the only way to know for sure how crooks might be using a stolen key, the cybercriminals who are reselling stolen or exposed AWS credentials for sex chats have started including programmatic checks in their code to ensure they aren’t using AWS keys that have prompt logging enabled.

“Enabling logging is actually a deterrent to these attackers because they are immediately checking to see if you have logging on,” Ahl said. “At least some of these guys will totally ignore those accounts, because they don’t want anyone to see what they’re doing.”

In a statement shared with KrebsOnSecurity, AWS said its services are operating securely, as designed, and that no customer action is needed. Here is their statement:

“AWS services are operating securely, as designed, and no customer action is needed. The researchers devised a testing scenario that deliberately disregarded security best practices to test what may happen in a very specific scenario. No customers were put at risk. To carry out this research, security researchers ignored fundamental security best practices and publicly shared an access key on the internet to observe what would happen.”

“AWS, nonetheless, quickly and automatically identified the exposure and notified the researchers, who opted not to take action. We then identified suspected compromised activity and took additional action to further restrict the account, which stopped this abuse. We recommend customers follow security best practices, such as protecting their access keys and avoiding the use of long-term keys to the extent possible. We thank Permiso Security for engaging AWS Security.”

AWS said customers can configure model invocation logging to collect Bedrock invocation logs, model input data, and model output data for all invocations in the AWS account used in Amazon Bedrock. Customers can also use CloudTrail to monitor Amazon Bedrock API calls.

The company said AWS customers also can use services such as GuardDuty to detect potential security concerns and Billing Alarms to provide notifications of abnormal billing activity. Finally, AWS Cost Explorer is intended to give customers a way to visualize and manage Bedrock costs and usage over time.

Anthropic told KrebsOnSecurity it is always working on novel techniques to make its models more resistant to jailbreaks.

“We remain committed to implementing strict policies and advanced techniques to protect users, as well as publishing our own research so that other AI developers can learn from it,” Anthropic said in an emailed statement. “We appreciate the research community’s efforts in highlighting potential vulnerabilities.”

Anthropic said it uses feedback from child safety experts at Thorn around signals often seen in child grooming to update its classifiers, enhance its usage policies, fine tune its models, and incorporate those signals into testing of future models.

Update: 5:01 p.m. ET: Chub has issued a statement saying they are only hosting the role-playing characters, and that the LLMs they use run on their own infrastructure.

“Our own LLMs run on our own infrastructure,” Chub wrote in an emailed statement. “Any individuals participating in such attacks can use any number of UIs that allow user-supplied keys to connect to third-party APIs. We do not participate in, enable or condone any illegal activity whatsoever.”

This Windows PowerShell Phish Has Scary Potential

Many GitHub users this week received a novel phishing email warning of critical security holes in their code. Those who clicked the link for details were asked to distinguish themselves from bots by pressing a combination of keyboard keys that causes Microsoft Windows to download password-stealing malware. While it’s unlikely that many programmers fell for this scam, it’s notable because less targeted versions of it are likely to be far more successful against the average Windows user.

A reader named Chris shared an email he received this week that spoofed GitHub’s security team and warned: “Hey there! We have detected a security vulnerability in your repository. Please contact us at https://github-scanner[.]com to get more information on how to fix this issue.”

Visiting that link generates a web page that asks the visitor to “Verify You Are Human” by solving an unusual CAPTCHA.

Clicking the “I’m not a robot” button generates a pop-up message asking the user to take three sequential steps to prove their humanity. Step 1 involves simultaneously pressing the keyboard key with the Windows icon and the letter “R,” which opens a Windows “Run” prompt that will execute any specified program that is already installed on the system.

Step 2 asks the user to press the “CTRL” key and the letter “V” at the same time, which pastes malicious code from the site’s virtual clipboard.

Step 3 — pressing the “Enter” key — causes Windows to launch a PowerShell command, and then fetch and execute a malicious file from github-scanner[.]com called “l6e.exe.”

PowerShell is a powerful, cross-platform automation tool built into Windows that is designed to make it simpler for administrators to automate tasks on a PC or across multiple computers on the same network.

According to an analysis at the malware scanning service Virustotal.com, the malicious file downloaded by the pasted text is called Lumma Stealer, and it’s designed to snarf any credentials stored on the victim’s PC.

This phishing campaign may not have fooled many programmers, who no doubt natively understand that pressing the Windows and “R” keys will open up a “Run” prompt, or that Ctrl-V will dump the contents of the clipboard.

But I bet the same approach would work just fine to trick some of my less tech-savvy friends and relatives into running malware on their PCs. I’d also bet none of these people have ever heard of PowerShell, let alone had occasion to intentionally launch a PowerShell terminal.

Given those realities, it would be nice if there were a simple way to disable or at least heavily restrict PowerShell for normal end users for whom it could become more of a liability.

However, Microsoft strongly advises against nixing PowerShell because some core system processes and tasks may not function properly without it. What’s more, doing so requires tinkering with sensitive settings in the Windows registry, which can be a dicey undertaking even for the learned.

Still, it wouldn’t hurt to share this article with the Windows users in your life who fit the less-savvy profile. Because this particular scam has a great deal of room for growth and creativity.