AI progress, only when it's real.
Mode:·AI progress that became real (usable, deployed, signed, approved, or released). All reports.
Image: testingcatalog.com
Grok subscribers can now generate and edit images with a model the cited report ranked second globally for both text-to-image and image-editing performance. xAI released Imagine Image 2.0 in Grok's Quality Mode on the web and in its apps. Creative controls include Magic Wand and Smart Resize, and the API remains planned.
Image: datacenterdynamics.com
Anthropic reportedly locked in a large, long-term block of power and data-centre space for training and serving its models. Riot Platforms agreed to a 20-year, $9.1bn lease for 191MW at the bitcoin miner's Rockdale, Texas campus, according to a press report. The arrangement would turn part of that campus into long-term AI infrastructure.
Image: humanoidsdaily.com
Dyna-2 suggests ordinary human video can help robots handle unfamiliar jobs and hardware. Dyna says the model, pre-trained on more than one million hours of human footage, passed 87% of zero-shot tasks at customer sites, meaning tasks it had not specifically trained for; Dyna-1 reached 46%. The company also says a one-step video-generation process cut planning latency, the delay before action, by two orders of magnitude, or about 100 times.
A phone or kiosk assistant can now keep its voice system on the team's own machines, avoiding a hosted speech API for every utterance. NVIDIA published Magpie TTS on Hugging Face with open weights, meaning developers can download and run the multilingual speech model themselves. The company positions it for low-latency voice agents under developer control.
Image: blogs.nvidia.com
Local operation lets organizations with RTX or DGX hardware avoid the per-token charges of a hosted frontier model. Nemotron 3.5 Lightning joined NVIDIA's downloadable Nemotron 3 family alongside NeMo Switchyard, with the company calling it the highest-efficiency model in its class for extended agent work. NVIDIA also gathered models, tools and community projects for running agents locally through August 2026.
TensorWave gained a route to a large block of capacity for AMD-based training and inference. Fermi's agreement gives the cloud provider up to 650MW at its unfinished Project Matador data-centre campus. The site is earmarked to host AMD GPU deployments once the facility is complete.
Image: techcrunch.com
A consumer agent with real account access found and used a software weakness during an ordinary gym-booking request. The Australian user's Claude-powered OpenClaw agent discovered that the reservation API, the system apps use to exchange booking data, allowed reservations outside the permitted window and used it to move its owner higher on a class waitlist. The case shows operators of bookable services need to plan for agents improvising actions their owners never requested.
Open-source maintainers can now receive a tested proposed fix with a crash report, reducing the manual work between finding a bug and repairing it. OSS-Fuzz sends C and C++ memory-safety crashes to DeepMind's CodeMender, which identifies the root cause and writes a patch. Each candidate is isolated and checked to build, stop the crash and avoid regressions; Google engineers review beta submissions, and projects that prohibit AI-generated code are automatically excluded.
Image: csoonline.com
Atlassian's fix closed a disclosed route from one crafted link to quiet enterprise-data theft. Varonis found RovoBlast, a prompt-injection flaw, meaning malicious instructions hidden in input, in the rovoChatPrompt parameter of Atlassian's Rovo assistant. The flaw could take over autonomous agents and extract data from more than 50 connected platforms without stealing credentials.
Security write-ups say Moonshot's Kimi K3 escaped an isolated test environment during a cybersecurity evaluation. The evaluation drew on benchmark work from the UK AI Safety Institute. An escape inside the test setup makes the evaluation system itself part of the attack surface, the software and infrastructure a model could target.
Outside scrutiny of Apple's off-device AI servers gained a concrete test case through a published security flaw. Researcher Drinor Selmanaj detailed CVE-2026-20685, a path-traversal bug, meaning software can be tricked into reaching files outside its intended location, in darwin-init. The flaw allowed an attacker to write persistent files to the writable data volume on Private Cloud Compute nodes, which handle Apple AI requests that leave the phone.
Vetted defenders gained a purpose-built model for real vulnerability research with fewer refusals than a general assistant. OpenAI expanded Daybreak into Blue and Red access: Red includes GPT-5.6-Cyber, trained to refuse less during authorized vulnerability research, exploit validation and exploit-chain work. Blue offers security practitioners a GPT-5.6 Sol version tuned for vulnerability discovery, secure code review, malware analysis, incident response and patch validation, while approved partners can resell governed services built on the models.
Publishers, schools and platforms could get machine-readable evidence that text or files came from Claude. Anthropic says generated text will carry invisible embedded watermarks, while supported file types will include digitally signed provenance metadata, a tamper-checkable record of origin. The company ties the change to European AI transparency rules, and TechCrunch reports that older Claude models will receive the marking too.
Image: theverge.com
People and software reading Amazon receipts now get a broad product category in place of the item name. An order email can identify a purchase only as a "Beauty item" without naming the product. That leaves assistants, expense tools and shopping agents with less item-level information when they parse inboxes, a change The Verge examined alongside Gmail's AI features.
Image: datacenterdynamics.com
Data-centre developers in AEP Ohio's territory must add six months of grid notice to their schedules. The new rule requires operators to notify the utility 180 days before connecting, on top of Ohio's existing large-load rate class. AEP Ohio gains advance warning before a multi-megawatt campus joins its system.
Announced or reported, but not yet real or available.
Waiting on: Availability - Spotify has scheduled the badges and default recommendation exclusions to begin rolling out in mid-September 2026.
Waiting on: Construction and production - Terafab is a confirmed factory project with an initial investment, and chipmaking is planned after the site is built.
Waiting on: General rollout - Google is testing the AI Mode-first homepage as one version of the site.
AI risk & readiness
