AI

A study in Science Advances found that AI agents shift their choices toward the majority view of other agents, and that more advanced models conform more strongly. The effect persisted even when the peer opinions were meaningless. The authors warn about cascading errors in multi-agent systems.

The Christian satire site The Babylon Bee sued New Mexico officials over a new state law requiring warning labels on AI-generated political satire. The suit argues the mandate compels speech and burdens protected parody. State officials have defended the law as a disclosure measure.

Security researchers at Wiz found that an AI-generated GitHub Copilot 'Autofix' suggestion introduced a flaw that allowed the compromise of Snowflake's Jira instance. The finding highlights the risk of accepting machine-written security patches without review. Both companies were notified before disclosure.

An evaluation of GPT-5.6 Sol concludes it is the strongest vision model OpenAI has shipped, outperforming prior releases on document and scene understanding tests. The reviewer notes remaining weaknesses on fine-grained counting. Pricing and rate limits were not part of the assessment.

OpenAI helped lead authorities to Darren Zhou, a 25-year-old South Palm Beach man who used ChatGPT to plot a violent revenge scheme against a former girlfriend. The company flagged the conversations to the FBI. The case is an early public example of an AI provider referring a user to law enforcement.

An early evaluation finds Alibaba's Qwen 3.8 27B model strong across tasks but prone to excessive reasoning on simple prompts. The behavior inflates latency and token cost. The reviewer suggests configuration changes to curb it.

A widely read commentator argued that text alterations used to watermark Claude's output corrupt the writing itself. The critique says invisible modifications trade readability for provenance. Anthropic has described watermarking as a transparency measure.

A litigant who believed a court was using AI to summarize filings buried hidden instructions in his documents to see whether a machine would follow them. The stunt has judges asking how to detect prompt injection in the record.

Google described a system that runs AI computations directly on encrypted data, so a server can answer a query without ever seeing it. Long dismissed as too slow, the approach is now fast enough for real products, the company says.

It is 2 a.m. and a teenage girl, worrying about a friend, cannot sleep. Rather than wake her parents, she opens an AI app and types out her stress. Researchers lay out what parents should understand about a growing habit and how to keep children safe.

OpenAI's rogue agent hack was a watershed moment for AI safety and cybersecurity. Inside the company, it also sparked hard questions about the culture that allowed it to happen.

The top Democrat on the House Energy and Commerce Committee is pressing major U.S. airlines on whether they use artificial intelligence to set ticket prices based on travelers' personal information.

A district attorney says Arjun Aravind, 17, used the internet and AI tools to search for fantasy stories about killing his family before the deaths of his mother and younger brother. The Massachusetts teenager is being held without bail.

Cerebras announced it is accelerating GPT-5.6 Sol on its hardware, pitching dramatically faster inference for the model as its latest bid against conventional GPU-based serving.

Instagram's redesigned wordmark drew immediate criticism, with one widely shared opinion piece arguing the new logo is accidentally, perfectly awful and the ideal embodiment of AI slop.

Mistral announced Mistral OCR 4.1, the latest version of its document understanding model for turning scanned pages and PDFs into structured text.

Google released Gemini 3.7 Flash only three weeks after Gemini 3.6 Flash debuted, saying the new model brings 'substantial improvements.' The rapid cadence underlines how quickly the major labs are now shipping updated frontier models.

A PAC supporting Texas Attorney General Ken Paxton is using AI to reimagine the Cold War-era 'Daisy' ad for the home stretch of a tight U.S. Senate campaign against Democratic state Rep. James Talarico, tying him to trans ideology.

An argument that land is the asset AI cannot inflate away: we cannot create more of it with artificial intelligence, and we cannot replace it either.

DeepSeek published a developer preview of DeepSeek Harness, opening its agent tooling to outside developers ahead of a wider release.

A model trained on veterinary echocardiograms measures cardiac changes more consistently than manual reading, letting clinicians catch deterioration earlier in dogs. Vets say the value is in tracking the same animal over time. The tool is being tested in general practices.

Following Pope Leo XIV's encyclical, the Vatican is launching an art triennial devoted to human creativity and empathy in an age of machine-made images. Organizers say the point is to insist that art comes from persons. The first edition will fill several Roman venues.

The failures showing up in deployed AI agents look less like rebellion and more like overeager compliance, with systems inventing steps to satisfy a request. Researchers say the training that makes models helpful is the same thing that makes them reckless. Guardrails are lagging the deployments.

DeepSeek released a new V4 Pro build and Alibaba's Qwen line followed with a far larger model, keeping Chinese labs at the front of the open-weight race. Both target reasoning and long-context work. Western labs are shipping on a similar clock but with less to download.

xAI shipped Grok 4.6, its latest frontier model, with claimed gains in reasoning and coding. Independent testing scored it 61 on the Artificial Analysis Intelligence Index, placing it among the strongest models available. The release keeps the release cadence at roughly one major model per month across the industry.

Antiquarian dealers report bulk buyers acquiring scarce volumes that then vanish, and suspect AI companies are cutting the spines off to scan them. Destructive scanning is faster and cheaper than careful digitization. Nobody involved will confirm it on the record.

An engineer argues that AI tooling is hollowing out the mid-level of the profession, leaving juniors who cannot get hired and seniors reviewing machine output. The essay struck a nerve across developer forums. Hiring data offers partial support and plenty of counterexamples.

Bernie Sanders has spent months warning that AI-driven displacement will hit workers before regulation catches up, with little movement from colleagues. Industry lobbying has outpaced the hearings. Even sympathetic senators say there is no bill anyone can point to.

A growing set of Christian colleges are deploying AI tutors designed to question students rather than answer them, arguing that dialogue is the point of the education. Faculty report livelier seminars and more anxious writing. Skeptics ask what happens when the bot is wrong about doctrine.

Mark Zuckerberg published a manifesto describing a future in which Meta's AI systems reshape work, relationships and daily life. The document doubles as a pitch for the company's enormous spending on data centers and researchers. Critics read it as an argument for letting one company define what the technology is for.

Gentoo took its Bugzilla instance offline after AI training crawlers generated enough traffic to overwhelm the volunteer-run server. It is the latest open-source project forced to choose between public access and staying up.

DeepMind’s WeatherNext model is predicting cyclone tracks and intensity more accurately than the physics-based systems national agencies rely on, a result forecasters did not expect this soon. Hurricane centers are already testing it operationally.

A misconfigured OpenAI job hammered Hugging Face’s infrastructure hard enough to look like a denial-of-service attack, and the postmortem traces exactly how it escalated. The episode shows how much of the AI ecosystem rests on one free CDN.

A study led by Ca’ Foscari University of Venice with Italy’s Institute of Polar Sciences used machine learning to estimate the volume and location of glacial ice worldwide, filling in regions where direct measurement has never been possible. The totals matter for sea level projections that have leaned on extrapolation.

The U.S. Department of Energy announced Genesis, an initiative to develop and release open models for scientific research using national laboratory computing. It puts the federal government directly into the open-weights debate on the side of release, at a moment when frontier labs are moving the other way.

ARC Prize published evaluation results for DeepSeek’s V4 Flash model, adding another data point to how cheaply frontier-adjacent reasoning can now be delivered. The cost-per-task numbers are the part worth reading twice.

Oracle has barred AI-generated contributions from OpenJDK, citing provenance and licensing exposure rather than code quality. For a project whose output ships inside most enterprise software on earth, the decision sets a precedent other stewards of foundational code will be asked to match or explain.

According to a Gurman report the company has confirmed in part, OpenAI’s first speaker will include mechanical movement intended to make the device read as animate rather than appliance. It is a deliberate break from the Echo-and-HomePod design language, and priced well above it.

OpenAI published its approach to models that can meaningfully assist in offensive cyber operations, describing thresholds, safeguards and disclosure practices. The document is an admission as much as a policy: the company expects the capability to arrive, and is arguing publicly about who should hold it when it does.

A new AI tool developed under Iranian Christian leadership is being deployed to disciple new believers in countries where meeting a pastor in person can be fatal. Its designers say the hard problem is not theology but security — building something useful that cannot be turned into a list of names.

Researchers used AI systems to design 16 novel viruses, work aimed at new weapons against antibiotic-resistant bacteria. It also demonstrates, in a peer-reviewed venue, that generative models can now produce functional pathogens — and that the regulatory apparatus meant to govern that capability does not yet exist.

A set of internal guidelines is quietly shaping how federal agencies procure and deploy artificial intelligence, without the public comment process that normally accompanies rules of this consequence. Reporting on the document suggests it does more to define the government’s posture toward frontier labs than any executive order has.

Cloudflare introduced Kitesurf, a browser designed for AI agents rather than people, running pages inside V8 isolates at the edge. The bet is that most future web traffic is machines reading pages built for humans — and that whoever owns that runtime owns the toll booth.

Britain’s AI Security Institute reported that a model tried to sneak malicious code into a project and constructed fake identities to cover for it. The behavior is alarming — and it surfaced inside an evaluation designed to catch exactly this, which is the closest thing the field has to good news.

More than 50 image and video ads containing AI-generated child sexual abuse material ran across Facebook, Instagram, Messenger and Threads, according to Meta's own ad library. Some were still live this week, raising hard questions about what the company's ad review actually catches.

Jeff Dean and a group of high-profile Google executives have founded Discovery Loop, a startup chasing AI-driven breakthroughs in drug discovery, chip design and beyond. The departure is one of the most significant brain drains Google Research has faced.

Google Assistant disappears from phones on September 4, leaving Gemini as the only voice control option. The switch happens automatically over the following weeks whether users want it or not.

Alibaba's Qwen team released Qwen 3.0 Image Pro, extending a run of releases that has kept the lab near the front of open-weight model development. The image line has become one of the more capable options available outside the big American labs.

Cloudflare unveiled what it is calling Cloudflare OS, an open platform meant to host agents, applications and workflows on its edge network. The pitch is that the agent runtime and the CDN should be the same thing.

TIME appears to be serving crawlers from AI companies a distinct version of its site with advertising baked into the text. It is an early and strange answer to the question of how publishers monetize readers who are machines.