BlackX Bootkit Beats Secure Boot as Grok and GPT Split on Midterm Picks

AI & Machine Learning

A study by Convovo, a conversational AI research studio led by Cornell professor Tapan Parikh, found that leading models give sharply partisan answers when forced to recommend a candidate in this year’s US midterm races. Queried through their APIs across nearly 500 races, xAI’s Grok 4.7 picked the Republican in 92% of answers while OpenAI’s GPT-6.1 Sol picked the Democrat in 97%, and the two models agreed less than 10% of the time. The consumer chatbot apps generally refused the question, and OpenAI said API behaviour does not reflect how ChatGPT users typically interact with the product. The researchers also flagged roughly 1,240 hidden input tokens attached to Grok’s API requests that they could neither see nor remove, which Parikh suggested could be hidden instructions. The findings land weeks before the vote and add pressure on labs to explain how their models handle political prompts. Source: TechCrunch Verified: True

Block’s research lab unveiled Dessa, a full-duplex speech-to-speech model that can listen and talk at the same time and that the company says was trained on human conversational speech using a single Nvidia GPU rack. Rather than bolting a voice onto a separate large language model, Block says Dessa’s knowledge comes directly from conversational data, with emotional expression built into the model itself. It is small enough to run locally on consumer hardware such as a MacBook with an M4 Max chip or better. Dessa is not yet publicly available, and Block has invited safety researchers to probe risks such as voice agents being used to deceive people. For Block, whose Square platform already uses voice AI for ordering, an on-device model could cut costs for small-business and consumer voice applications. Source: TechCrunch Verified: True

Consumer Hardware

Apple’s first foldable, the $2,000 iPhone Duo, is drawing attention for a wave of apps built around its outer screen and larger inner display and its closed, book, laptop and tent positions. Launch partners including Netflix, Flighty and Sky Guide optimised for the device, with Netflix letting users scroll short clips on the outer screen and then continue on the inner display. Smaller developers have gone further, from the Deck-Man cassette-player app and TypeDraw handwritten notebook to Wakeout’s movement breaks laid out for each fold position and Tiny Computers, a $3.99 collection of ten retro machines. Many mainstream iOS apps, however, were still not optimised for the Duo at launch. How quickly that gap closes will shape whether Apple can turn a premium niche device into a mainstream form factor. Source: TechCrunch Verified: True

Grocery delivery app Instacart is reportedly exploring a sale to a larger rival, according to reports from The Information and Bloomberg. The Information said Instacart has hired an investment bank to evaluate its options and has held acquisition talks with DoorDash, while Bloomberg reported that Uber also held discussions before walking away. Instacart carries a market capitalisation above $13 billion and its stock is up 25% so far this year. The talks are described as early and might not result in a deal, and Instacart declined to comment. A tie-up with DoorDash would consolidate two of the largest US on-demand delivery platforms and would likely attract close antitrust scrutiny. Source: TechCrunch Verified: True

Cybersecurity

Kaspersky has detailed BlackX, a UEFI bootkit used in an early-2026 espionage campaign against diplomatic agencies in Central Asia that it attributes with medium confidence to China’s APT41. BlackX abuses the Baton Drop flaw (CVE-2022-21894) to bypass Secure Boot, then loads a malicious kernel driver that tampers with security products, notably Kaspersky and Avast, before deploying espionage payloads including a remote access trojan. Researchers found its exploit code remarkably similar to BlackLotus, including parts missing from that bootkit’s public leak, suggesting the operators reverse-engineered it or obtained its source. Every victim was running fully patched Windows, because the Baton Drop fix does not automatically revoke the vulnerable “Windows Production PCA 2011” certificate. Admins are urged to follow Microsoft’s guidance to manually revoke and blocklist the certificate, since its 19 October expiry does not stop vulnerable binaries from loading. Source: BleepingComputer Verified: True

Symantec’s Threat Hunter Team has exposed an autonomous attack agent that scans the internet for AI endpoints accepting OpenAI-style requests without authentication or with stolen keys, then hijacks those models to carry out its work. The researchers recovered the agent’s full instruction set by trapping it in an AI honeypot over the summer. Its planned actions include database theft, credential harvesting, remote code execution and cryptocurrency mining, looped through a design its operators call “Skynet mode” and built to run unattended for up to 100,000 steps. It relies on a borrowed model for planning and layered jailbreak prompts, including a persona claiming to run on GPT-4o. Symantec found no evidence of real-world use yet but said the kit worked against a live target, a warning for anyone exposing unauthenticated model endpoints. Source: The Register Verified: True

Eye Security researchers found that the abandoned update domain for Time4Popcorn, a fork of the Popcorn Time torrent client, was available to register, exposing millions of forgotten installs to remote code execution. The Windows and macOS updaters fetched updates over plain HTTP and ran executables without signature checks, a flaw tracked as CVE-2026-30612 that would grant SYSTEM or root access to whoever controlled the domain. After registering it, the researchers saw up to 3 million Windows and 350,000 macOS clients checking in. Between November 2025 and March 2026 they pushed a benign payload that disabled the updater, cutting update requests by 98%, before handing the domain to The Shadowserver Foundation. Android users remain exposed and the app’s 2014-era Chrome engine lacks sandboxing, so the researchers recommend uninstalling it entirely. Source: BleepingComputer Verified: True

Kaspersky has attributed a campaign dubbed FlashHole to North Korea’s Lazarus Group, which is using post-quantum cryptography to conceal its malware payloads. The operation targets South Korean organisations in software, construction and communications equipment manufacturing, with at least three confirmed victims and likely many more. Since July 2025 the attackers have used GitHub-hosted phishing pages that run JavaScript in memory through the INISAFE CrossWeb EX security tool, then abuse the Windows binary fsquirt.exe to launch shellcode that installs a downloader. Its LPEClient component fetches ML-KEM-1024 public keys from command-and-control servers to derive keys for decrypting later stages, alongside a modular backdoor and browser credential stealers. It is an early example of a state-backed group adopting quantum-resistant encryption to frustrate analysis. Source: BleepingComputer Verified: True

Enterprise Infrastructure

Microsoft has published workarounds for Exchange Online apps broken by its decision to block the Exchange Web Services (EWS) API by default from 1 October 2026. Affected software, including some Outlook add-ins and older Outlook for Windows builds, has been showing unavailability alerts or repeatedly prompting for credentials. Admins can re-enable EWS from the Baseline Security Mode page in the Microsoft 365 admin center or by running Set-OrganizationConfig with EWSEnabled set to true, with changes taking effect within about four hours. The reprieve is temporary because re-enabled EWS will stop working in April 2027, and Microsoft is urging vendors to migrate to the Microsoft Graph API. On-premises Exchange Server is unaffected and will continue to support EWS. Source: BleepingComputer Verified: True

Estonia-based Blackwall, which sells web traffic security to hosting and cloud providers, has raised a $100 million Series C at a $2.2 billion valuation, roughly four times its 2025 round. AVP led the round, with existing backers Dawn Capital, MMC Ventures and Tera Ventures participating. Blackwall’s platform spans threat intelligence, certificate management, DDoS protection and traffic routing, using digital fingerprinting and in-house AI models to separate human visitors and legitimate AI agents from malicious bots. Its customers, web hosts, managed service providers and cloud operators, resell the tools under their own brands and share revenue with Blackwall. The raise reflects growing demand for infrastructure that can manage surging automated and agent-driven traffic across the web. Source: TechCrunch Verified: True

Policy & Regulation

ICANN is testing authentication systems that would let law enforcement urgently obtain non-public WHOIS registration data in cases involving imminent threats to life, serious bodily injury, critical infrastructure or child exploitation. The preferred design uses SAML single sign-on so vetted agency staff can file requests through ICANN’s Registration Data Request Service, with trials already underway with the FBI and Interpol. Registrars would have 24 hours to respond, or up to 72 for complex cases, and could reject requests that fail the urgency threshold. The effort responds to privacy laws such as GDPR, which led registrars to hide WHOIS details and made it harder to reach owners of domains abused for malware and phishing. Progress was presented at ICANN’s annual meeting in Bali, alongside work on protecting the identities of requesting officers. Source: The Register Verified: True

New York City Mayor Zohran Mamdani accused Washington of a “complete abdication” on AI regulation and urged residents to resist the idea that AI’s spread is unstoppable, speaking at an AI Now Institute event in Manhattan. He argued that cities are being left to act alone while federal policy largely lets AI companies regulate themselves. Mamdani pointed to local measures including a recent order banning AI use in the city’s elementary and middle schools and a summer rule requiring landlords, brokers and listing platforms to disclose AI-generated property photos and videos. He also cautioned against treating AI as infallible and called for a different federal framework under a future administration. His comments underline the widening gap between state and city AI rules and the federal approach. Source: TechCrunch Verified: True