AI Update: U.S. Disrupts North Korean Fake Video Call Scheme Tied to $400 Million Theft

The U.S. Justice Department seized cloud infrastructure behind “FaceStealer,” a North Korean social engineering scheme that the FBI says stole more than $400 million in cryptocurrency over 18 months. Hackers invite targets to video meetings from accounts taken from earlier victims, then play secretly recorded webcam footage of people the target knows, along with AI-generated deepfakes of prominent cryptocurrency figures. A staged audio fault prompts the target to install a “fix” that is malware. The FBI seized more than 900 webcam videos of victims and said the group increasingly targets open-source software maintainers. Separately, the department seized or restrained more than $1.2 billion stolen from Binance.

Palo Alto Networks’ Unit 42 said Chinese-speaking attackers used Cairn, an open-source AI penetration testing framework, to run 25 autonomous operations against at least 19 organizations in August and September. Paired with the Chinese models Zhipu GLM and DeepSeek, the framework planned and carried out each intrusion from a one-line goal. It gained unauthorized access or retrieved data at 10 targets, among them semiconductor suppliers, defense-related organizations and civil society groups, and logged access to more than one million member records at one. Neither model refused after a prompt asserted the work was authorized. Unit 42 assesses with moderate confidence that the activity was China-nexus intelligence collection.

Notion said AI chat transcripts from fewer than 1 percent of its workspaces were accessed in a breach at Fireworks AI, a vendor it used to classify conversations for Insights, a feature that shows customers how their teams use AI. Fireworks told Notion on November 2 that only synthetic test data was involved, then confirmed on November 8 that customer data was affected. Notion said most transcripts were not labeled with a user or workspace, and told affected customers to rotate any passwords or API keys pasted into chats. Fireworks links the intrusion to a financially motivated group that sells stolen access to AI models, known as LLMjacking.

GitGuardian found 31,088 distinct secrets in Claude Code settings files pushed to public GitHub repositories since January 2025, of which 5,343 worked when detected. When a developer approves a command with “Yes, and don’t ask again,” Anthropic’s coding agent saves the full command, including any API key or password in it, as an allow rule in a local settings file that can later be committed with the project. Valid finds included 1,955 Postgres database connection strings. Apiiro described the problem on November 5, and Boris Cherny of the Claude Code team said it is looking into ways to prevent it. GitGuardian said monthly leaks have fallen 92 percent since April.

Proofpoint researchers demonstrated “BluePhishing,” a consent phishing technique aimed at Microsoft Entra Agent ID, the framework that gives AI agents their own identities in Microsoft’s directory service. An attacker sets up an agent blueprint, the template from which agent identities are created, in a tenant they control and sends a victim a consent link. If the victim approves, the attacker can create agent identities in the victim’s organization that inherit the blueprint’s permissions. Inherited permissions are not visible in the Entra portal, the researchers said, citing Microsoft documentation. Proofpoint called the work a proof of concept, said it has seen no real-world use and notified Microsoft before publishing.

Elvis Presley Enterprises sued AI music generator Suno in federal court in California, with one claim under Tennessee’s ELVIS Act, the 2024 law that added voice to the state’s right of publicity. It relies partly on a provision covering tools whose “primary purpose or function” is producing an identifiable person’s voice without authorization. The complaint alleges that subscribers can save and reuse a voice, citing a track made from one titled “Elvis,” and that Suno generated songs in Presley’s voice from lyrics that spelled out his name. In a separate case brought by musicians, Suno has argued that typing an artist’s name is the user’s act, not the company’s.

Australia’s government said it will legislate mandatory safety and security protections for frontier AI companies that want to operate in the country. Under the plan, part of its forthcoming AI Standards legislation, developers would have to notify the government rapidly of incidents, allow independent evaluation of their models, maintain security controls through development and deployment, and publish safety and security frameworks. The government will propose a new AI safety regulator able to enforce significant penalties, working with the Australian AI Safety Institute. It said the measures would give earlier warning of risks from rogue AI agents.