NerdNews

Meta's new policy document, the Frontier AI Framework, outlines scenarios where the company may stop development of AI systems deemed too risky, including those that could aid in cybersecurity, chemical, and biological attacks.

Bluesky is a decentralized social app that has gained over 30 million users. It was conceptualized by former Twitter CEO Jack Dorsey and is developed using an open-source framework called the AT Protocol. Bluesky has features similar to X, including a Twitter-like user interface, algorithmic choice, and community-specific moderation. The app is free and plans to sustain itself through paid services, including custom domains and a subscription service called Bluesky+. Bluesky has been growing rapidly, but its growth has slowed down recently.

OpenAI has launched a deep research agent for ChatGPT, which can find, analyze, and synthesize hundreds of online sources to create a comprehensive report. The agent uses a version of the company's upcoming o3 model and can take anywhere from 5 to 30 minutes to complete its work. It has shown an accuracy of 26.6% on the Humanity's Last Exam evaluation, outperforming other models like GPT-4o and Grok-2.

OpenAI's models have shown rapid progress in making human-level persuasive arguments, with the new o3-mini model ranking as more persuasive than humans in about 82 percent of random comparisons. However, the company notes that this is still short of 'superhuman' performance and is taking steps to mitigate potential risks of AI-powered persuasion.

SoftBank is investing $3 billion annually to develop and market AI agents with OpenAI, aiming to automate over 100 million workflows among its companies, including tasks like generating financial reports and managing customer inquiries.

Tana, a startup founded by ex-Googlers, has raised $25M for its AI-powered knowledge graph that aims to simplify work by organizing and automating tasks. The tool can transcribe conversations, turn them into action items, and integrate with various apps. With a 160K+ waitlist, Tana is poised to revolutionize productivity.

A US senator has proposed a bill that would impose penalties of up to 20 years in prison or $1 million in fines for individuals who download or invest in Chinese AI models, and up to $100 million in fines for US enterprises that collaborate with Chinese companies on AI research. The bill aims to restrict AI-related trade and collaboration with China.

Researchers at the Okinawa Institute of Science and Technology built a brain-inspired AI model and put it in a robot to help it understand the world. The AI learned concepts behind words by interacting with its environment, demonstrating compositionality and generalizing knowledge to new tasks.

OpenAI's deep research is an autonomous research assistant that compiles full-blown research reports. It works by asking follow-up questions to clarify the user's query, then methodically goes through sources to synthesize an answer. The tool can take anywhere from 1-30 minutes to deliver a report, which can range from a few thousand to 16,000 words. While it's not perfect and can be slow or repetitive at times, deep research has the potential to revolutionize the way we access and interact with information.

Large language models are vulnerable to prompt injection and stealing attacks, which can be used to access business data, gain personal advantages, or exploit tools. To defend against these attacks, developers can add instructions in the system prompt, use adversarial detectors, or fine-tune their models. However, there is no reliable solution yet that completely prevents these attacks, and active defense is necessary.

Google has fixed a high-severity zero-day vulnerability in the Android kernel that has been exploited in attacks. The vulnerability, tracked as CVE-2024-53104, is a privilege escalation flaw in the USB Video Class driver. Additionally, the February 2025 Android security updates patch 48 vulnerabilities, including a critical security flaw in Qualcomm's WLAN component.

DeepSeek, a Chinese AI company, has developed an open-source AI model that can rival the performance of Big Tech companies' models, but at a significantly lower cost. This has sent shockwaves through the tech industry, with many companies reevaluating their AI investment strategies. The article discusses the implications of DeepSeek's achievement and how it may change the AI landscape.

Anthropic has released a new system of Constitutional Classifiers to filter out jailbreaks in its Claude AI model. The system has withstood over 3,000 hours of bug bounty attacks and is now open for public testing. The classifiers use a 'constitution' of natural language rules to define permitted and disallowed content, and can detect and block jailbreak attempts with a high degree of accuracy.

A cloud security firm found a publicly accessible database belonging to DeepSeek, containing over 1 million instances of chat history, backend data, and sensitive information. The database was completely open and unauthenticated, allowing for full control and privilege escalation.

Elon Musk is taking control of the US government, pushing out the head of the FAA, accessing Treasury systems, and locking out civil servants. His goal appears to be a coup over the administrative state, with the support of President Trump. Musk's actions are causing chaos, including the shutdown of USAID and the potential disruption of government payments.

PowerSchool, a US edtech giant, has suffered a massive data breach affecting millions of students. The breach, which occurred in December 2024, involved unauthorized access to the company's school information system, PowerSchool SIS. The company has been criticized for its lack of transparency regarding the breach, including the number of affected schools and students, the types of data stolen, and the amount paid to the hacker. PowerSchool has hired cybersecurity firm CrowdStrike to investigate the breach, but the results of the investigation have not been made public.

The US Department of Homeland Security's Inspector General has launched an audit of the Transportation Security Administration's use of facial recognition technology at US airports. The audit will evaluate the effectiveness and privacy concerns of the technology, which has been criticized by lawmakers and privacy advocates. The TSA has been using facial recognition to screen passengers at US airports for several years, with plans to expand to 430 airports in the next decade.

A 22-year-old Canadian man, Andean Medjedovic, has been charged with stealing approximately $65 million by exploiting vulnerabilities in DeFi protocols KyberSwap and Indexed Finance. He allegedly used deceptive trades to drain millions of dollars from the protocols, and then laundered the proceeds through various means. Medjedovic faces multiple charges, including wire fraud, unauthorized damage to a protected computer, and money laundering.

US President Donald Trump plans to impose tariffs on foreign semiconductors, including those used by Nvidia. Despite a meeting with Nvidia's CEO Jensen Huang, Trump reaffirmed his plans to impose tariffs, which could lead to higher costs for electronics in the US. The tariffs are part of Trump's ongoing trade war with China and other countries.

A US judge has denied Apple's request to delay the upcoming Google trial, which could potentially end Apple's $20 billion annual deal with Google for default search on iPhones. The trial will determine remedies for Google's monopolistic practices, and Apple had sought to intervene, but the judge ruled that Apple's motion did not meet the requirements for a stay.