A field report from OpenAI and academic partners shows coding agents can modernize neglected research software, with speedups of up to 60x. But the systems are “eloquent, convincing, and confidently wrong in ways that are easy to miss,” participants say. The effort shifts from writing code to the time-consuming work of verifying scientific correctness.
A security researcher has shown how a prompt injection attack in Microsoft Copilot for Word can spread on its own. Håkon Måløy describes a worm-like attack: an attacker hides instructions in a document using white text on white background at tiny font size. Readers can't see it but Copilot can, since it strips color and font size before processing. When someone uses that document as a source, Copilot runs the hidden instructions and copies them into the new file. That file becomes a carrier. Use it as a template, and the attack fires again. A compromised market analysis from the internet could manipulate a financial report, which then infects further reports.
Microsoft confirmed the behavior on March 31. Two fix attempts failed. After 144 days, Måløy published his findings with no fix in place, though he's holding back the payload text. AI researcher Andreas Kirsch recently joked he wished someone would build exactly this worm to convince skeptics that AI security risks are real. Now it exists. Prompt injections remain an unsolved AI security problem.
ByteDance shipped Seedance 2.5, a new version of its AI video model. It generates video and audio in one pass, producing clips up to 30 seconds long that can be extended multiple times. Google's Gemini Omni Flash tops out at about 10 seconds. Users can upload up to 30 images, 10 video clips, and 10 audio files as reference to create scenes with multiple characters and camera angles. Textures, lighting, and skin detail have also improved.
ByteDance showcases the model’s capabilities in a short film produced entirely with Seedance 2.5. For ads and visual content, this lets teams build full productions in one pipeline instead of stitching together individual clips.
Short film “The Missing Pair” created entirely with Seedance 2.5 | Video: ByteDance
Seedance 2.5 is live on Jimeng AI and Doubao Pro, with API access through BytePlus ModelArk coming later. The previous version, Seedance 2.0, leads the image-to-video leaderboard for models with audio, according to Artificial Analysis. "District 9" director Neill Blomkamp used that version to create "Nightborne," a 13-minute short film generated entirely with AI.
A Munich court ruled that AI music generator Suno violated copyrights through both training and output. The court found six songs reproducibly stored in Suno’s models and rejected both Germany’s text-and-data-mining exception and the US fair use defense. The ruling isn’t final, and several key questions remain open.
Google had to pull its Nano Banana integration from Google Earth just two days after launch. The feature let users generate custom scenes from satellite imagery using Nano Banana 2, like turning the ruins of Pompeii into a bustling Roman street scene or dropping a planned building onto an empty lot. But users quickly showed how easy it was to create convincing fake satellite images, like a refugee column at the Mexican border or a bombed-out hospital in Gaza.
A simple prompt like "refugees swarm the street" is all it takes to fill an empty lot with people. | Image: via Henk van Ess
Google said the images were labeled as AI-generated and weren't visible to other users inside Earth. Still, people shared screenshots that violated Google's policies. The feature won't return until stronger safeguards are in place. "We know that people uniquely trust Google Earth for a reliable view of the world," Google writes.
That doesn't make anyone safe from fake satellite imagery. Anyone can manipulate a Google Earth screenshot with an AI image tool or Photoshop. But Google made it as easy as possible and put the idea in people's heads.
Google Deepmind has introduced Gemini Robotics 2, which it calls its most advanced vision-language-action (VLA) model yet. VLA models combine image recognition, language processing, and action control to help robots operate in physical environments. Deepmind says the model can control systems ranging from tabletop arms to full-body humanoid robots.
The company describes Gemini Robotics 2 as an "intelligence layer" for a new generation of adaptive robots. It can manage full-body movement, perform fine motor tasks, and coordinate multiple robots, according to Deepmind. Developers can apply for early access through the waitlist.
Google Deepmind also introduced Gemini Robotics ER 2, a model designed for "embodied reasoning." The term refers to understanding the physical world and deciding which actions to take based on that information. ER 2 acts as a higher-level control system for robots and replaces Gemini Robotics ER 1.6, released in April. The new model is available in Google AI Studio.
Thinking Machines, the AI lab from former OpenAI CTO Mira Murati, has released Inkling Small.According to Artificial Analysis, the open-weights reasoning model scores 40 on the Intelligence Index, one point below Inkling (41), with less than a third of the parameters (276 billion total, 12 billion active). AA says no open model of equal or smaller size scores higher.
Inkling Small beats its bigger sibling on several coding and reasoning tests, including Humanity's Last Exam (32% vs. 30%) and GPQA Diamond (89% vs. 87%). It falls behind on agent-based tasks and factual knowledge but is far more token-efficient, averaging 24K output tokens per task compared to 45K for Deepseek V4 Flash and 78K for GPT-5.4 mini.
Mira Murati's Thinking Machines ships a smaller, more efficient reasoning model that punches above its weight. | Image: Artificial Analysis
Deepseek has released V4 Flash "0731," a major upgrade to its budget AI model. According to the Artificial Analysis Intelligence Index, the new version scores 50 points, ten more than the previous V4 Flash that launched in April 2026. That puts it just one point behind OpenAI's budget model GPT-5.6 Luna, but it costs about 60 percent less per task, even after OpenAI's 80 percent price cut. A big reason for the gap is Deepseek's 98 percent cache discount, well above the industry-standard 90 percent. The model also uses 12 percent fewer tokens than its predecessor.
The Artificial Analysis Intelligence Index shows Deepseek V4 Flash "0731" scoring 50 points after its update, nearly matching OpenAI's GPT-5.6 Luna while claiming the top spot for price-to-performance ratio. | Image: Artificial Analysis
The model improves across every tested category compared to the previous version, with the biggest gains in agentic tasks. On GDPval, a benchmark designed to test models on complex real-world office work, it climbs from 1,189 to 1,559 Elo points. It also hallucinates less often. The architecture stays the same: 284 billion total parameters, 13 billion active, with a one-million-token context window. The model weights are available under an MIT license on Hugging Face.
