Binance Traces $6.3B Heist to Fake Job Interview as Anthropic Nears RSI

AI & Machine Learning

Anthropic’s revised S-1/A filing warns that the company may be close to recursive self-improvement (RSI), the point at which AI systems become good enough at training their successors that capability gains accelerate sharply. The filing admits Anthropic has previously used AI to build models whose capability jumps it did not recognise at the time, and says future models may soon meet its Responsible Scaling Policy definition of “dramatically accelerated progress attributable to AI.” No current model has been designated as crossing that threshold, and Anthropic says it does not expect to safely let Claude design successive frontier models without meaningful human oversight. It is weighing policy updates to keep humans in the loop, though sceptics note that IPO risk disclosures can double as marketing. Source: TechCrunch Verified: True

Anthropic also published FlashProve, an open-source system for “provable inference” that cryptographically proves which model produced a given output. Labs would publish a model fingerprint that can be matched against outputs without revealing weights, letting users detect silent swaps to cheaper or different models, even when the lab serving the model is the one doing the swapping. Verification is designed to be light enough to run on a local device, and the specification and code are on GitHub. Anthropic calls it intermediate progress and says there are no current plans to deploy it at Anthropic or elsewhere, but it targets a growing trust gap in hosted AI. Source: TechCrunch Verified: True

Inference platform Fal raised $1.3 billion in a round led by Bessemer Venture Partners, with Nvidia, Sound Ventures and T. Rowe Price participating, doubling its valuation to $16 billion from $8 billion in April. Announced together, the Series E and F rounds total $1.65 billion, and the San Francisco company, which hosts and runs video, image and audio generation models, has passed $800 million in annualised revenue. The raise shows how much money is moving into serving already-trained models rather than training new ones. Rivals such as Baseten, Fireworks AI and Modal are growing quickly too, but high compute costs keep gross margins across the category thin. Source: TechCrunch Verified: True

Consumer Hardware

Apple told developers it is tightening App Store rules so minors can’t easily reach dating apps or AI companions offering sexual or romantic content. Any app with dating or romantic-relationship features must now carry an 18+ rating, and from November 16 developers must add new content descriptors so age ratings reflect what their apps actually do. Apps that share a user’s real-world location with another person or group must also disclose it. The change builds on Apple’s global age-assurance rollout and new parental controls, and it lands squarely on AI companion apps such as Character.AI and Replika, which have drawn heavy scrutiny over their appeal to younger users. Source: TechCrunch Verified: True

Netflix launched an “AI Discover” tab as a testbed for new ways to pick something to watch, starting with mood-based recommendations. Viewers choose curated themes such as “something light” or “scare me,” which also take into account viewing history, available time and what’s trending, then refine the results with a “Tune Your Recommendations” control that saves or excludes titles in real time. The feature is rolling out first to English-language US users on supported TVs, shown as a “Discover” button marked “New” in the top navigation, with wider availability promised in the coming weeks. It’s Netflix’s most visible move yet to put generative-AI-style discovery on the living-room screen. Source: TechCrunch Verified: True

Cybersecurity

Binance says the record theft of about $6.34 billion on October 29 began with a fake job interview, attributing the attack to North Korea’s TraderTraitor group (also tracked as Lazarus, Jade Sleet and APT38). In September an engineer was approached on a job-hunting platform and interviewed by impostors posing as an HR representative and a company CEO, then given a coding assignment whose repository carried undetected macOS malware disguised as a pyenv component. The attackers used that foothold and an internal AI tool to compromise two employees on the wallet infrastructure team by October 20, stole a container deployment token to reach the wallet signing service, and abused Ethereum’s EIP-7702 to keep draining funds after withdrawals were paused. About $2.85 billion has since been frozen or intercepted, and Binance says it had replenished 98% of affected token assets by October 31. Source: BleepingComputer Verified: True

CrowdStrike says a China-nexus group it tracks as Vapor Panda used ARTEX, an open-source multi-agent AI penetration-testing framework, to attack more than 200 organisations in 14 countries starting October 17. Targets included government ministries, telecoms and pharmaceutical, semiconductor, defence, nuclear and financial firms; at least nine were compromised and at least 12 more hit with password spraying and credential scanning. The operators coordinated agents with DeepSeek and an unnamed commercial model, gave instructions in Chinese, and told the agents to avoid Falcon-protected Windows hosts, with one intrusion consuming over 1.4 billion input tokens. Separately, the Dutch Institute for Vulnerability Disclosure said an autonomous ARTEX agent reached root on its network in 12 minutes in September by chaining two zero-days. Source: BleepingComputer Verified: True

Barracuda Networks urged customers to patch two flaws that together can give attackers root on its appliances. CVE-2026-116606 is a critical authentication weakness in the Email Security Gateway that can let attackers change an appliance’s global configuration, while CVE-2026-111560 lets users with admin rights inject OS commands that run as root across ESG, Load Balancer ADC, Message Archiver, WAF Control Center, Web Application Firewall and Web Security Gateway. Chained, they allow an attacker to bypass authorisation and take over a device, though both require access to the admin interface, so Barracuda advises limiting that to trusted internal networks. Patches are out for all affected products and no active exploitation has been reported, but ESG has been a prized target before through CVE-2023-2868 and CVE-2023-7102. Source: BleepingComputer Verified: True

Enterprise Infrastructure

AI data centre builders stuck in years-long queues for grid connections and new turbines are turning to repurposed jet engines, rebuilt to burn natural gas, for on-site power. The International Energy Agency puts the wait to connect 100 megawatts in Northern Virginia at about seven years, while a jet-engine setup can reportedly get a site running in 24 months. xAI has installed eight such units in Mississippi for its Colossus 2 data centre, and Crusoe has ordered 13 converted engines for its Abilene, Texas campus, where OpenAI also relies on gas turbines. The converted engines are less efficient and dirtier than conventional gas turbines, so experts see them as a stopgap, even as the Trump administration encourages data centres to generate their own power. Source: TechCrunch Verified: True

AMD is on track to overtake Intel as the largest x86 chip vendor by revenue after a third quarter in which sales rose 45% year over year to $13.4 billion. Data Center revenue nearly doubled, up 97% to $8.5 billion, while client and gaming fell 8% to $3.7 billion even as PC processor sales grew 7% to $2.9 billion on Ryzen mobile and enterprise demand. AMD guided to roughly $17 billion next quarter against Intel’s nearly flat $15 billion to $16 billion forecast, which would put AMD ahead. CEO Lisa Su called 2027 “a very strong momentum year,” citing data centre and AI demand and a supply chain able to meet it. Source: The Register Verified: True

Policy & Regulation

The New York Times has sued Brave in federal court in Manhattan, and the Daily News, Chicago Tribune and Orlando Sentinel filed a parallel suit the same day. The Times alleges Brave scraped its copyrighted journalism with an unidentified crawler publishers cannot block, built an index of more than 30 billion pages, and sold that content to AI companies it names as Anthropic, Perplexity, Mistral, xAI and Cohere. The complaint also targets a Brave browser feature that bypasses paywalls and the company’s practice of blocking publishers’ ads while selling its own. The Times is seeking an injunction and damages, opening a new front in the AI copyright fight aimed at the search-index layer rather than the model makers themselves. Source: TechCrunch Verified: True

Google says that next year it will switch off AI tools for K-12 students by default, covering Gemini, NotebookLM, Gemini in Classroom and AI in Google Search, unless school administrators explicitly turn them on. Because Google says search and AI are now deeply integrated, schools that block AI entirely will also lose Google Search, and students will see a message saying it is unavailable. The change, which Google says followed consultation with schools, teachers, researchers and policymakers, arrived alongside Guided Learning, a tutor-style AI tool. Critics argue the all-or-nothing choice uses Google’s dominance of classroom Chromebooks to push AI adoption on schools that rely on search for research. Source: TechCrunch Verified: True