AI‑enhanced malware automates victim profiling for faster credential theft

Volta-autodoxer adds AI to infostealer malware

Cisco Talos researchers disclosed a new strain of infostealer malware called Volta-Autodoxer that incorporates large language model (LLM) functionality to profile victims and prioritize high‑value data. The malware automatically gathers credentials, tokens and payment information before sending a serialized dossier to an LLM for analysis¹.

Volta-Autodoxer collects passwords, browser‑saved credentials, Microsoft OAuth tokens, cookies, payment cards and crypto‑wallet seed phrases, as well as active sessions on Discord, Telegram, Steam and Roblox. After the initial sweep it performs a deeper scan of central accounts such as Gmail, extracting receipts, order history and location data from services like UberEats¹.

Infostealer malware has long been used to harvest login details that fuel large data breaches, including incidents involving Snowflake, AT&T, Ticketmaster and Disney. Cisco Talos flagged the tool using its new CAIRN framework, which detects hacking tools that embed AI capabilities, though the researchers have not observed active deployments in the wild¹.

“We’ve realized that we need to connect these agents to a lab to get the maximum value from them”

Apple CEO on AI strategy

Apple CEO John Ternus said the company will use artificial intelligence in upcoming devices, but it will not release AI features just for their own sake. He became CEO in September and addressed the topic in a Wired interview with Steven Levy. Ternus emphasized that the goal is to create products people love, regardless of whether AI is involved².

Ternus noted that Apple has been perceived as lagging because Siri’s more powerful version only arrived this summer. He argued that when AI is applied correctly, users should not need to consider the underlying technology. He also said that AI could eventually make tasks like app creation easier, but product creation remains a human‑driven activity².

The CEO’s comments come as the industry debates whether AI agents might replace traditional apps. Ternus’ stance suggests Apple will continue to prioritize user experience over showcasing AI for its own novelty².

Pika H3 Flex video‑audio model architecture

Pika Labs released Pika H3 Flex, a post‑trained H3 model for professional creators, on 29 October 2026. The model generates video with synchronized stereo audio in six sampling steps. It supports up to 12 reference inputs and can produce native 480p or 768p video locally, with higher resolutions via the hosted API³.

The architecture adds two task adapters: FL2VA for text and endpoint‑frame conditioning, and Ref2VA for multimodal references. It builds on the MiniMax‑H3 base model and processes prompts and media through an encoder before producing video and audio latents. The inference pipeline then decodes these latents and combines them into an MP4 file, automatically downloading the pinned base components and selected adapter³.

Pika H3 Flex integrates both modalities, allowing creators to control motion and sound together, which simplifies multi‑shot storytelling and motion‑graphics workflows³.

OpenAI lowers funding target to $20 billion

OpenAI announced that it has reduced the size of its upcoming financing round to a $20 billion target after earlier reports of a $30 billion raise. The revision follows a valuation drop from $1.4 trillion to $1.25 trillion and comes ahead of a planned initial public offering. Investors from the United Arab Emirates will lead the round, with a possible retail tranche for wealthy individuals⁴.

The company has already secured $11 billion in commitments, according to a Wall Street Journal follow‑up, but the overall response from large investors has been described as “tepid”. The first close of the round is expected to be below the $20 billion goal, and the total may only be reached through the additional retail offering. Analysts suggest that the softer fundraising environment is linked to Wall Street’s caution over heavy AI infrastructure spending and a broader sell‑off in tech stocks⁴.

AI agent performs brain injection on live mouse

Biotech founder Gabriel Ocana Santero connected an AI agent to a robotic arm, injector and camera to perform a brain injection on a living mouse. The experiment was carried out by his company Exin Therapeutics in San Francisco and was reported on 31 October 2026. The AI chose a therapy from 19 options, set injection coordinates and controlled the needle during the procedure⁵.

The setup included a motorized device to immobilize the mouse, a camera for visual feedback and a robotic injector that the AI could move. The team first tried Anthropic’s Claude Opus 5.5, which refused to guide the needle, citing internal safety warnings. They switched to Claude Sonnet 5, which completed the injection as instructed⁵.

AI models used in research are typically trained on published papers, so they miss real‑time observations and failed experiments. By giving an AI direct control of lab hardware, researchers hope to make models more reliable and useful for experimental design. Santero’s test shows both the promise and the current safety limits of such integration⁵.