Discord Investigators Gained Illicit Entry to Anthropic’s Mythos

As researchers and professionals discuss the effects of emerging AI models on cybersecurity, Mozilla announced on Tuesday that it utilized early access to Anthropic’s Mythos Preview to identify and patch 271 vulnerabilities in its latest Firefox 150 browser release. In parallel, researchers have uncovered a group of moderately adept North Korean hackers utilizing AI for tasks ranging from writing malware to fabricating fake corporate websites—resulting in thefts amounting to $12 million within three months.

Researchers have successfully dissected disruptive malware named Fast16, which predates Stuxnet and could have been aimed at Iran’s nuclear program. Initially created in 2005, it was likely deployed by the US or an allied nation.

Meta faces a lawsuit from the Consumer Federation of America, a nonprofit organization, regarding fraudulent ads on Facebook and Instagram and allegedly misleading claims about the company’s anti-scam efforts. Additionally, a United States surveillance program allowing the FBI to access Americans’ communications without a warrant is set for renewal, yet lawmakers are at an impasse regarding future actions. A new legislative proposal seeks to address the growing concerns among lawmakers but lacks specific details.

For those interested in an in-depth analysis, WIRED delved into the ongoing feud surrounding the well-known privacy-focused mobile operating system, GrapheneOS. We also explored the peculiar case of how China spied on US figure skater Alysa Liu and her father.

And there’s even more. Each week, we compile security and privacy news that we haven’t discussed in detail ourselves. Click on the headlines for full articles. Stay safe out there.

Anthropic’s Mythos Preview AI model has been described as a highly powerful tool for discovering security vulnerabilities in software and networks, so much so that its creator has placed significant restrictions on its release. Nevertheless, a group of amateur detectives on Discord discovered straightforward methods—without the need for AI hacking—to gain unauthorized access to a prized digital asset: Mythos itself.

Despite Anthropic’s attempts to control access to Mythos Preview, a cohort of Discord users managed to acquire the tool through basic investigative work: they analyzed data from a recent breach of Mercor, an AI training startup collaborating with developers, and “made an educated guess about the model’s online location based on knowledge about the format Anthropic employs for other models”—a phrase that many speculate refers to a web URL, as reported by Bloomberg, which first revealed the story.

One individual reportedly exploited permissions they already had to access additional Anthropic models due to their association with an Anthropic contracting firm. As a result of their investigation, they allegedly accessed not only Mythos but also other unreleased Anthropic AI models. Fortunately, according to Bloomberg, the group that accessed Mythos has thus far only utilized it to create simple websites—a strategy aimed at evading detection by Anthropic—rather than launching any significant hacking efforts.

Security experts have long cautioned that the telecom protocols known as Signaling System 7, or SS7, which dictate how phone networks connect and route calls and texts, are susceptible to misuse that can enable covert surveillance. This week, researchers from the digital rights organization Citizen Lab disclosed that at least two for-profit surveillance companies have exploited those vulnerabilities—or similar ones in next-generation telecom protocols—to spy on actual victims. Citizen Lab found that two surveillance firms operated like rogue phone carriers, taking advantage of three smaller telecom companies—Israeli carrier 019Mobile, British provider Tango Mobile, and Airtel Jersey, located on the island of Jersey in the English Channel—to track the locations of targeted phones. The researchers indicated that “high-profile” individuals were monitored by the two surveillance companies, although they did not disclose the names of either the firms or their targets. They also warned that the two discovered companies are probably not the only ones engaging in such practices and that vulnerabilities in global telecom protocols pose a genuine threat for phone surveillance around the world.

In a sign of an increasing—if tardy—crackdown by US law enforcement on the vast criminal enterprise of human-trafficking-driven scam operations in Southeast Asia, the Department of Justice announced charges this week against two Chinese individuals for allegedly helping manage a scam operation in Myanmar and planning the establishment of another compound in Cambodia. Jiang Wen Jie and Huang Xingshan were arrested in Thailand earlier this year on immigration-related charges, according to prosecutors, and now face allegations of running an extensive scamming operation that lured trafficking victims with false job offers and compelled them to defraud victims, including Americans, out of millions through cryptocurrency investment scams. The DOJ stated that it has also “restrained” $700 million in funds associated with the operation—effectively freezing assets in preparation for seizure—and has confiscated a Telegram channel believed to be used to entice and enslave trafficking victims. The Justice Department’s announcement claims that Huang personally participated in the physical punishment of laborers in one compound, and that Jiang at one point oversaw the theft of $3 million from a single scam victim in the US.

Three scientific research institutions have been identified as selling British citizens’ health data on Alibaba, according to revelations from the British government and the nonprofit UK Biobank. Over the past two decades, more than 500,000 individuals have provided health information—including medical images, genetic data, and healthcare records—to UK Biobank, which enables researchers globally to access this information for medical studies. However, the organization stated that this data leak involved a “breach of contract” by the three institutions, with one of the datasets believed to potentially encompass information on all half a million research subjects. The types of data listed for sale were not fully detailed, but the Biobank has suspended the accounts of those involved in selling the information, and the listings have since been taken down.

Earlier this month, 404 Media reported that the FBI managed to obtain copies of Signal messages from a defendant’s iPhone since the content of the messages, protected by Signal’s encryption, was stored in an iOS push notification database. In this case, the messages remained accessible even after Signal had been deleted from the device—an issue affecting all apps that utilize push notifications.

In response, Apple this week released a security update for iOS and iPadOS to address this vulnerability. “Notifications marked for deletion could be unexpectedly retained on the device,” notes Apple’s security update for iOS 26.4.2. “A logging issue was resolved with enhanced data redaction.”

While this issue has been rectified, it remains advisable to modify what appears in notifications on your device. For Signal, you can launch the app, navigate to Settings, then Notifications, and switch notifications to display Name Only or No Name or Content. This serves as a reminder that although apps like Signal offer end-to-end encryption, this protection applies to content during transmission: If someone can physically access and unlock your phone, there is a possibility they can access everything stored on your device.