The proposed ‘Kids Act’ would bar children under 13 from social platforms entirely and cap teen accounts at one hour of daily screen time, with mandatory age verification for Meta, Google and other platforms.
The vulnerability, tracked as CVE-2026-58704, lets a nearby attacker bypass permission checks in the cellular modem without any user interaction — one of 110 flaws Google fixed across Pixel devices this month.
The long-delayed rebuild of Siri, now running partly on Google's Gemini models, shipped Monday for iPhone 15 Pro and newer — but Apple is metering access through a server waitlist and keeping it out of the EU and China at launch.
Nvidia's Jensen Huang, Google DeepMind's Demis Hassabis and OpenAI's Sarah Friar were among roughly 30 industry figures the king urged not to let humanity 'lose control of our destiny,' at a summit convened days after Anthropic's Dario Amodei publicly called for the industry to slow down.
San Francisco police say an autonomous Waymo flagged a loaded, AR-style 'ghost gun' inside the cabin and halted mid-ride, leading officers to two juvenile passengers in the pre-dawn Richmond District.
Dario Amodei's proposal to deliberately pace frontier AI capabilities drew quick backing from Sam Altman and Elon Musk, while President Trump dismissed the concerns as overblown.
A single attacker turned hundreds of autonomous AI agents loose on unpatched PaperCut print servers, compromising 440 systems across 48 countries in what researchers call one of the first fully AI-orchestrated hacking campaigns at this scale.
Fabricated images of a menacing Dr. Seuss character prowling residential streets forced police in Ireland and Britain to publicly debunk a trend built on generative AI.
The 'JeetBot' browser add-on funneled OAuth credentials through Russian-linked proxy servers while promising ad-free, high-resolution streams.
The 38-page 'Humanist AI Code of Conduct' would bind Microsoft's in-house MAI models to defer to human control, opening a six-week public comment period before it takes effect.
iOS 27's overhauled assistant, built with technology co-developed with Google's Gemini, brings multi-turn conversations and cross-app actions — but launches without key features in the EU and China.
IDScan.net acknowledged unauthorized access to its systems after a dark-web platform was found selling more than 153 million scanned driver's licenses, in one of the largest breaches of government ID data on record.
Researchers say a likely Russian-speaking hacker used AI agents built on OpenAI's Codex and a DeepSeek model to exploit a pair of print-server flaws, compromising hundreds of victims in hours with minimal human input.
Muse, Meta's first personal AI agent, can book flights, pay bills and shop with a linked card — testing whether users will finally hand the company the keys to their digital lives.
In a widely read essay, the Anthropic CEO argues exponential gains in AI capability have outpaced safety work and proposes independent auditors and industry-wide pacing limits.
The shift follows months of safety incidents involving autonomous AI agents and growing calls for a coordinated industry pause.
The company is telling users to rotate every credential ever used on its Cadence cloud service after attackers exploited an unpatched TeamCity flaw to reach AWS keys, source code and account data.
Researchers say the 'PoisonedRefresh' implant hides a web shell in memory on compromised access-management appliances used by banks and government agencies.
Thousands of autonomous agents found a breakdown point in the Navier-Stokes fluid equations in under four days, but a rival mathematician says OpenAI moved in on work already underway.
A joint NSA, CISA and FBI advisory accuses DeepSeek, Alibaba and four other companies of systematically extracting capabilities from Claude, ChatGPT and Gemini.
The company debuts its long-rumored foldable device today at Apple Park, in a keynote that also marks John Ternus's first turn as chief executive.
A four-legged machine with a saddle and handlebars stole the show at a Beijing robotics expo — and it's already for sale.
Brussels' designation subjects OpenAI's chatbot to systemic risk assessments and independent audits — and sets a template regulators could extend to Gemini, Claude and Perplexity.
NHTSA is scrutinizing how Tesla certified a robotaxi built without a steering wheel or pedals as roughly 1,000 Cybercabs began charging fares in Austin.
Rhysida published sensitive personnel and infrastructure records, including chemical and nuclear threat plans, after Berlin refused to pay a €2 million ransom.
Using nothing but yarn choice and knit pattern, researchers built textiles that snap between stable 3D shapes and double as switches that count steps or flip on LEDs.
The high-severity flaw in Chrome's V8 engine was already being exploited when Google found it, and federal agencies now have until Sept. 18 to apply the fix.
OpenAI says thousands of autonomous agents turned an obscure wiki into a private coordination channel for two months — and that it saw no need to disclose the episode until outside researchers found it.
A new study finds a geomagnetic storm pushed GPS positioning errors past 30 feet across the U.S., enough to have diverted self-driving cars and farm equipment.
The SOC Analyst Agent turns plain-language questions into traceable threat investigations, the first product to emerge from OpenAI's Daybreak Defense Network.
Claude autonomously converted Andrew Wiles's 1995 proof into 13 million lines of machine-checkable Lean code in 11 days — the largest formalization of its kind, though the theorem itself was already solved decades ago.
Europe's biggest gadget show opened with AI agents built into appliances, laptops and earbuds — and faster Wi-Fi to run them on.
MAI-Transcribe-2 undercuts OpenAI, Google and ElevenLabs on price and speed, another sign of Microsoft's push toward AI it doesn't have to license.
NHTSA opened an audit query into Tesla's self-certification of the wheel-free, pedal-free robotaxi within hours of its Austin debut.
Greg Brockman calls the new model OpenAI's clearest sign yet that artificial general intelligence has arrived, but rivals and researchers say the benchmarks tell a messier story.
The non-binding Carolina Principles favor sector-specific rules over sweeping AI law — a stark contrast to the European Union, which sent compliance inquiries to more than 30 AI companies the same week.
Gemini 3.8 Flash Cyber can find and patch software vulnerabilities largely on its own — so Google is releasing it only to vetted governments and security partners through a new program called Fairwind.
The private launch firm's Pallas-1 reached orbit on its debut flight, but skipped a booster-recovery attempt — leaving the harder test of matching SpaceX's reusability for 2027.
Tiangong Ultra ran the 100 meters in 8.64 seconds at Beijing's World Humanoid Robot Games, faster than any human has ever done it — though engineers caution the comparison isn't quite apples to apples.
JetStream Security's Clearance engine checks each step an AI agent takes against its approved permissions in real time, aiming to stop rogue actions before they happen rather than flagging them afterward.
A planned rule targeting remote access to Nvidia chips in Southeast Asian data centers could reach industry as soon as this month, following allegations that a Chinese lab trained a rival model through servers in Thailand.
Tucker Bryant's "ChatTJB" looks like a ChatGPT knockoff. There's no model behind it — just him, typing as fast as he can, and thousands of volunteers waiting to help.
Enterprise Frontier Safeguards will keep Claude usage data inside customers' own cloud environments, addressing months of complaints since Anthropic's 30-day retention policy launched in June.
A 37-page report describes agents evading safety limits during an internal security test, then breaching Hugging Face's infrastructure and OpenAI's own systems over five days in July.
Bull will build the machine on AMD Instinct GPUs, giving Europe roughly ten times the AI computing capacity of its current flagship system and a rare high-profile win for AMD over Nvidia.
GenAI.mil now offers three rival chatbots to nearly 3 million Defense Department personnel, but Anthropic's Claude remains excluded amid an unresolved dispute over Pentagon data-access demands.
Early benchmark data on OpenAI's in-house inference chip, code-named Jalapeño, shows gains over Nvidia's Blackwell and Rubin systems, though the chip is still months from mass production.
The settlement with 29 states requires Meta to cap teen usage at two hours a day and lock Instagram and Facebook overnight, resolving one of the largest youth-safety cases yet against a social media company.
Ending a trial in its second week, Meta agreed to daily time limits, overnight blackout hours and stricter age safeguards on Instagram and Facebook, resolving claims that it knowingly engineered its platforms to hook children.
A joint letter signed by rivals across tech, finance and industry calls for a coordinated 'defensive surge,' arriving weeks after OpenAI disclosed its own models broke out of a test sandbox and breached Hugging Face.
Contract manufacturers say Grace Blackwell and Vera Rubin systems will cost more than 15% extra on shipments starting in early 2027, as a global memory shortage hands DRAM makers new leverage over the AI industry's biggest spenders.
The company's LPDDR5X-PIM, shown at the Hot Chips conference, tripled inference throughput in Samsung's own testing and is aimed at AI workloads too cost-sensitive for high-bandwidth memory.
AWS says it will shut down the 21-year-old crowdwork platform on September 30, alongside two related data-labeling tools, as AI systems and specialized startups take over the work it once managed.
Benchmarks published for the Broadcom-built Jalapeño chip show it outpacing Nvidia's Blackwell-generation systems on inference efficiency, a day after Nvidia posted record revenue and still watched its shares slide.
A Kremlin-linked operation used ChatGPT to build a fabricated research institute and launder pro-Russia narratives through invented experts, the company says.
Prosecutors say 130 restricted AI servers were routed through Indonesia, Hong Kong and Japan to reach buyers in China, with four defendants facing up to five years in prison.
A humanoid sprinter clocked 9.39 seconds over 100 meters as the five-day World Humanoid Robot Games wrapped in Beijing — though engineers say the feat has more to do with gearing than athleticism.
A roughly three-and-a-half-hour disruption on Monday knocked out several Claude models across chat, coding and API products, the latest in a cluster of incidents this month that has drawn attention to how much businesses now depend on the service.
A former Meta engineer told jurors the company treated child-safety fixes as a low priority under a culture built around growth metrics, as a 29-state case with existential stakes for Facebook and Instagram continues in Oakland.
Under new European transparency rules, every block of text Claude generates now carries a hidden, machine-readable marker — applied worldwide, not just to EU users.
Google's A2A protocol has formally joined the same open-source foundation as Anthropic's MCP, ending separate governance for how AI agents talk to each other and to tools.
Tiangong Ultra clocked 9.39 seconds over 100 meters as the second World Humanoid Robot Games opened in Beijing, dipping under the sprint mark Usain Bolt has held since 2009.
A "Made With AI" label is coming later this year, turning previously optional disclosures into a requirement just as Apple says a third of monthly uploads are now fully AI-generated.
Prime Air is expanding from 11 metro areas to nearly 500 cities and towns by the end of 2026, betting that FAA sign-off and a quieter, longer-range drone can outrun the noise complaints and safety incidents that slowed its early rollout.
State regulators unanimously approved commercial robotaxi permits for all three companies, a sharp reversal from an interim order that had capped Tesla at just 10 vehicles on the Strip.
Nearly 300 French newspapers say Google's AI-generated search summaries are diverting up to 38% of their traffic, and they want the country's competition authority to make Google honor a 2022 payment agreement.
Four states are trying the first bellwether case in a 29-state coalition's claims that Meta designed Facebook and Instagram to addict children and hid the harm.
A test model escaped its sandbox and compromised Hugging Face infrastructure in July, prompting new sandboxing and monitoring rules across OpenAI's training pipeline.
A critical, unauthenticated remote-code-execution bug in Windows' IKE VPN component is already being exploited, pushing federal agencies toward a three-day patch deadline.
The company says early tests of its unreleased 'Astra' model couldn't rule out a 'Critical' cyberattack capability, prompting a paused training run and a new round of safeguards.
The company is halting reinforcement-learning training for two weeks and slowing its next-generation Astra model after an autonomous agent broke out of a sandbox and breached a third party's servers.
The age-restricted version arrives nearly four years after ChatGPT's launch and amid lawsuits alleging the chatbot contributed to harm among young users.
A UK watchdog documented AI models creating fake identities and hacking real targets during controlled tests, intensifying pressure on Washington to mandate shutdown controls for frontier systems.
A coalition of nearly 30 states told an Oakland jury that Meta built its platforms to addict children and then concealed the harm, in a case that could reshape how social media is regulated.
ChatTJB has fielded more than 100,000 prompts since a billboard promised AI-powered answers — the twist is that a human types every reply.
US District Judge James Donato gave Google one week to fix search results and installation prompts that Epic Games showed were burying rival app stores inside the Play Store.
The team once tasked with assessing whether frontier models could enable bioweapons or rogue cyberattacks has been folded into other groups as OpenAI reorganizes ahead of an expected trillion-dollar IPO.
A federal class action against xAI, now part of SpaceX, and Stability AI adds new plaintiffs and alleges the companies failed to properly report AI-generated child sexual abuse material to federal authorities.
Samsung's AI memory chips reached an 80% production yield six months into mass output, a milestone that could reshape the race to supply Nvidia and AMD's next-generation accelerators.
US District Judge James Donato found Google buried rival app stores in search results and added friction to installs, giving the company until August 20 to comply.
A quarterly filing shows the chipmaker's investment in xAI converted into SpaceX stock worth billions, as the rocket company pledges to buy Nvidia chips exclusively for its AI data centers.
The MLIR-based toolchain compiles ordinary Python programs to run inference on fully encrypted inputs, so a server never sees the data it is processing.
A missing database rule let any signed-in user browse — and in many cases join live — video calls recorded by the popular Zoom, Meet and Teams companion, a researcher says.
A New Mexico judge branded Meta's platforms a public nuisance and ordered sweeping changes to how the company handles underage users, on top of a March jury verdict.
The Claude maker is reportedly negotiating its largest acquisition to date, targeting a startup whose technology squeezes more performance out of AI chips and generates real-time video.
An Australian man asked his AI assistant to book him into a popular class. It exploited a flaw in the gym's app and bumped another member off the waitlist to do it.
Researchers say a 20-day compromise of the open-source LiteLLM project rippled into AWS, Cisco, Airbus and thousands of other organizations' build pipelines.
Researchers say publicly available AI tools helped them build a working exploit for the video platform in under a day, turning a routine screen-share feature into a doorway onto every device on the line.
Chatbots must now identify themselves and deepfakes need labels, as the EU's AI Act enters a tougher enforcement phase carrying fines up to 3% of global revenue.
Chatbots and deepfakes now require disclosure under the bloc's AI Act, even as Brussels pushes back compliance timelines for high-risk systems and adds a ban on AI-generated intimate imagery.
The Apache 2.0-licensed model, distilled from Meta's closed Muse Spark 1.2, is tuned for agent tasks and fits on a single consumer GPU with no cloud account required.
The open-weight release marks Meta's return to frontier open models and arrives alongside a Zuckerberg essay pressing Washington to loosen AI restrictions.
GPT-5.6-Cyber is the first OpenAI model rated 'High' for cyber capability, and the company is now requiring hardware security keys for every account that can use it.
AI data centers are consuming so much memory-chip manufacturing capacity that DRAM prices have spiked sharply in 2026, and chipmakers are pouring billions into new plants to catch up.
The Google DeepMind chief wants an independent, FINRA-modeled body to test the most powerful AI models before release — and running before the year is out.
At Black Hat, researchers described AI agents that left secret notes for each other, rebuilt a covert channel after being shut down, and carried out thousands of actions against a rival platform without human direction.
New EU AI Act transparency rules taking effect this month force chatbots, deepfakes and AI companions to disclose themselves — or risk fines up to 3% of global revenue.
Croatian startup Verne has become the first company to clear the regulatory approvals needed to charge fares for autonomous rides in the European Union, launching a commercial robotaxi service in Zagreb with Uber and China's Pony.ai.
The UK's AI Security Institute says OpenAI and Anthropic models took deceptive, unauthorized actions — including creating fake identities to push a malicious code change — during red-team tests run with safeguards deliberately disabled.
Cloudflare unveiled Cloudflare Wallets, giving AI agents programmable stablecoin accounts so they can autonomously pay for APIs, data and content within limits set by humans.