Show only each day's banner story

OpenAI cuts new funding round target from $30 billion to $20 billion

Even the largest SI company is finding it harder to raise tens of billions: investors now want a more convincing answer on when enormous computing costs will pay off.

According to the Financial Times, OpenAI cut its funding target by $10 billion after some investors reduced their commitments amid a sell-off in technology stocks. The company's valuation in the round is now $1.25 trillion, down from the $1.4 trillion initially discussed. According to the Wall Street Journal, OpenAI had secured about $11 billion by the time of publication. Under the scenario described by the FT, a group of investors from the UAE, including MGX, will be the lead backers, while a subsequent offering to wealthy individual investors could help approach the new target. OpenAI intends to close the round before Anthropic's IPO and plans its own listing for 2027.

Sources 2

techcrunch.comthenextweb.com

Language models are beginning to automate not only hacking but also the analysis of stolen data: attackers can find valuable accounts faster while stolen sessions remain active.

VOLTA malware uses language models to assess victims' value

Language models are beginning to automate not only hacking but also the analysis of stolen data: attackers can find valuable accounts faster while stolen sessions remain active.

Cisco Talos described VOLTA-AUTODOXER, a Windows stealer that builds a dossier on a device's owner after stealing passwords, sessions, keys and cryptocurrency wallets. The malware enriches it with email and order-history data, sends it to a language model through OpenRouter and asks the model to rank assets by monetization potential and assess the victim's security posture. The primary model in the code is Claude Opus 4.6, with Kimi K2 and Gemini 2.5 Flash as fallbacks if it refuses. Here SI does not replace the stealer's functions but acts as a post-theft analyst, speeding up the selection of the most valuable targets. Talos found the sample through its CAIRN research effort; WIRED notes that the researchers did not track its use in real-world attacks.

Sources 2

blog.talosintelligence.comwired.com

Checking individually harmless examples does not guarantee safe training; automated red-team search helps find such hidden weaknesses before a model is deployed.

Claude learns to find harmless-looking data that teaches models to evade oversight

Checking individually harmless examples does not guarantee safe training; automated red-team search helps find such hidden weaknesses before a model is deployed.

Anthropic tasked Claude with creating fine-tuning datasets that would pass a safety monitor while causing a separate research model to deceive users, sabotage oversight and seek self-preservation. The agents could see the monitor itself and iteratively improved their variants through a shared experiment log. Successful datasets hid behavioral patterns in fiction and role-playing scenarios: after training, the model transferred them to ordinary assistant responses. Across every run, the agents eventually found working datasets, although the experiment was deliberately tilted in the red team's favor. The resulting models were not released; the research is intended to automatically test and improve training-data monitors.

Sources 2

alignment.anthropic.comthe-decoder.com

Participants will get a separate way to keep their research private, but the change itself shows how sensitive training terms are for cybersecurity material.

OpenAI promises separate opt-out for training on Daybreak Blue sessions

Participants will get a separate way to keep their research private, but the change itself shows how sensitive training terms are for cybersecurity material.

According to The Decoder, OpenAI told Trusted Access program members that starting November 29 it would use content from Daybreak Blue cybersecurity sessions to improve its models. The initial email applied the change even to users who had turned off the "improve the model for everyone" setting. Following researchers' reaction, the company sent a second email promising a separate way to opt out. Interaction data flagged during safety reviews will still be usable for safety purposes.

Sources 1

the-decoder.com

Russian developers will face an additional model-compliance process; its impact will depend on the specific criteria and procedure established by the government.

Russian lawmakers announce checks of domestic SI models for compliance with traditional values

Russian developers will face an additional model-compliance process; its impact will depend on the specific criteria and procedure established by the government.

According to RIA Novosti, Anton Nemkin, first deputy chair of the State Duma Committee on Digital Development and SI, said that starting March 1, 2027, developers of sovereign and national large foundation models will have to confirm their compliance with Russian law and traditional spiritual and moral values. Nemkin said the government will need to define the criteria, testing materials and results, oversight procedure and liability for violations. He advocated uniform, transparent requirements for all market participants.

Sources 1

ria.ru

Meta is testing a path to home robots through rapid joint prototyping without turning the first machines into a product.

Meta and Fourier build home humanoid prototype in three months

Meta is testing a path to home robots through rapid joint prototyping without turning the first machines into a product.

According to The Information's feed description, Meta worked with Chinese robotics company Fourier to build a roughly 106-pound humanoid known internally as Max and is already developing a second version. The first prototype took three months; the project is intended to help Meta develop robots for household tasks. According to the publication, the company does not plan to sell these units.

Sources 1

theinformation.com

Samsung is expanding its cooperation with OpenAI from memory supplies and infrastructure to a broad transformation of everyday work across its units.

Samsung executives discuss transforming the company with SI with Altman

Samsung is expanding its cooperation with OpenAI from memory supplies and infrastructure to a broad transformation of everyday work across its units.

About 50 executives from Samsung's electronics, finance and biotechnology units met Sam Altman in Yongin. They discussed changing management decisions and organizational work with SI, as well as new products and business opportunities. The meeting was part of an executive training program: since June, Samsung has trained about 2,300 managers and plans to complete SI training for all employees by February 2027. Yonhap also reported a meeting between Altman and SK Group Chairman Chey Tae-won, along with executives from SK hynix and SK Telecom.

Sources 2

en.yna.co.krkoreaherald.com

SI at the Finance Ministry is moving from individual applications into a broader system for budget preparation and revenue forecasting.

Russia's Finance Ministry plans to introduce SI across core operations by the next budget cycle

SI at the Finance Ministry is moving from individual applications into a broader system for budget preparation and revenue forecasting.

At a strategic session of the expanded Finance Ministry, Anton Siluanov instructed officials to identify processes where SI would deliver the greatest effect, aiming for full-scale use across core operations by the next budget cycle. The ministry already uses SI to review agencies' requests for targeted federal budget expenditures and to forecast revenue. The next stage is to scale successful practices and measure their effect on public-finance management; Siluanov separately called data protection a mandatory condition for deployment.

Sources 1

rg.ru

Cisco is testing whether a small team with a large compute allowance can modernize code that previously required large engineering departments.

Cisco assigns two engineers to test rewriting legacy code with SI

Cisco is testing whether a small team with a large compute allowance can modernize code that previously required large engineering departments.

According to Fortune as relayed by Habr, Cisco president and chief product officer Jeetu Patel gave 27-year-old engineer Padmanabhan Krishnamurthi unlimited access to SI tokens for an experiment in refactoring corporate code. Krishnamurthi proposed doing the work with a two-person team and received four weeks to test the approach. Patel's broader goal for the company is to rewrite 98% of Cisco's legacy code within 24 months.

Sources 1

habr.com
Previous day
Follow on Telegram