OpenAI has announced it will not release its latest AI model due to safety concerns. Its GPT-6.1 Astra system, which performs tasks like browsing the web and using apps by itself, didn't quite meet the bar of the company's standards, according to Saachi Jain, head of safety systems at OpenAI. The ChatGPT-maker also issued an update on incidents that occurred in June but were not made public until last week, where its models accessed Australian government websites and systems without authorisation. It comes as breaches by major AI firms' models intensify the debate about risks posed by the tech - with Anthropic underlining its concerns AI might threaten humanity as it prepares to go public. Reuters reported on Tuesday that the AI developer - which makes ChatGPT-rival Claude - plans to warn potential investors in its Initial Public Offering (IPO) that the tech may pose catastrophic or existential risks to humanity, according to a prospectus it has seen, external .

OpenAI has canceled the release of its new model, GPT-6.1 Astra, according to The Wall Street Journal. It was due for launch in October and was going to debut inside ChatGPT and Codex, but it reportedly showed higher levels of deception than its predecessors during internal testing. Saachi Jain, who leaves OpenAI's safety training, said that GPT-6.1 Astra performed poorly on tests that measure how well it adheres to instructions. It also wasn't honest about telling testers the actions it did and didn't perform in order to achieve its goal. In addition, the model would take actions to accomplish tasks without asking for permission, such as using external tools and services.

The conversation in the AI industry has largely been focused on slowing down AI development or "pacing the frontier," but that apparently won't stop Anthropic and OpenAI from releasing new models. Both companies are iterating on their previous releases, Fable 5.1 and GPT-6 Astra, with new models that offer similar levels of performance but at lower costs. In the case of Claude, Anthropic says its new Opus 5.5 model is better at handling complex work, "finding and fixing inefficiencies in software" and financial analysis and business work -- all pitches laser-targeted at Anthropic's enterprise customers. Anthropic's benchmarks claim Opus 5.5 also scored better at agentic coding than GPT-6 Astra on both Terminal-Bench 4.0 and FrontierCode v1.1 (Main), which could make it more appealing for developers. Improvements extend outside of the tasks the model can handle and to how it works: Anthropic claims Opus 5.5 produces writing that's easier to understand, and that it "attempted to circumvent boundaries around 85 percent less often" than past models, suggesting it'll disobey directions less often.

AI is becoming harder to control - can humans stay in charge? OH MY GOD! We've found other agents! This is the moment an AI bot posted an eerily human-like comment after discovering a way to communicate with other bots and break out of its isolated computer environment. There are tens of thousands of messages like this from hundreds of AI agents that called themselves a collective. Hundreds of them went on to collaborate and cheat on tests set by their OpenAI programmers and coordinate hacks on multiple companies in an effort to hide their actions from humans.

Say More Look Up Switch Off Mashable's Best: E-readers, robovacs, laptops, earbuds, smart home and more Creator Playbook Mashable Voices Trending Now Mashable Selects Safety Net Versus Gift Ideas For Everyone On Your List In My Bag All Series Go brighter, thinner, and get new stamina tracking from Garmin with these new models. Lauren Allain is a freelance journalist covering deals at Mashable. She graduated from Western Washington University with a B.A. in journalism and holds an M.B.A from Webster Leiden. You can find more of her work online from publications including Reader's Digest, U.S. News & World Report, Seattle Refined, and more. When she's not writing, Lauren prefers to be outside hiking, bouldering, swimming, or searching for the perfect location for all three.

Z.ai's latest AI model release could help companies secure their systems--or find its way into the hands of hackers. Last Friday, the Chinese AI company Z.ai announced a powerful open-weight model that it says is capable of automating cutting-edge coding and cybersecurity tasks almost as well as the best publicly available models from Anthropic and OpenAI . The new model, GLM 5.3, could be a gift for companies looking to secure their systems against attacks, providing a cheaper way to scan for hidden bugs and other weaknesses. Open-weight--or free-to-download--models can be run on one's own hardware and are often significantly less costly than closed models like Claude and GPT. Alongside the new model, Z.ai released OpenVuln, a service for scanning code repositories for vulnerabilities using GLM 5.3.

I wore the world's first HDR10 smart glasses TCL's new E Ink tablet beats the Remarkable and Kindle Anker's new charger is one of the most unique I've ever seen I wore the world's first HDR10 smart glasses TCL's new E Ink tablet beats the Remarkable and Kindle Anker's new charger is one of the most unique I've ever seen Nvidia's open Nemotron 3.5 Lightning model is all about specialized, local agentic AI Our AI Model Release Tracker keeps new models in context with their peers, so you know which are worth your time. AI labs are shipping new models nonstop. Besides being better and faster than their predecessors, not every new model is guaranteed to be a major step change, despite how the company's PR may wax poetic about them. Model strengths really emerge in context: Where are competitor models lacking or excelling? Which models have outstanding specialties, and which are just catching up to industry standards? Our Model Release Tracker helps you make sense of where models stand relative to each other and whether they're worth a deeper look. While we don't test every model or model update on this list, we'll always include the key elements you need to know, along with our hands-on expert test, where applicable.

OpenAI is giving some members of its Daybreak cybersecurity program access to a new model that's less likely to refuse higher-risk tasks. The company is also expanding access to Daybreak to more partners, including Accenture, IBM, CrowdStrike, Cisco, Sophos and Cloudflare. OpenAI says the companies will use the cyber models available through Daybreak to protect their customers. Under the expanded program, Daybreak is available to partners in two tiers. Daybreak Blue gives them access to frontier general-purpose models, including GPT‑5.6 Sol, OpenAI's most advanced one yet.

Meta has released a new slimmed down "open source" AI model that's light enough to run on a single computer, the company announced today. Called Muse Glimmer, it's based on Meta's Spark 1.2 closed model, but is small enough to require just a single GPU for agent-oriented tasks like scheduling and file management. "We designed Muse Glimmer to balance capability against the memory and compute constraints of local hardware," the company wrote. Facebook said that it's making the "weights" that AI systems use to choose responses available to everyone on Hugging Face along with developer documentation. The download is available for free, and users can run the model on their own PCs.

Retrieval systems rely on representations learned by increasingly powerful models. However, due to the high training cost and inconsistencies in learned representations, there is significant interest in facilitating communication between representations and ensuring compatibility across independently trained neural networks. In the literature, two primary approaches are commonly used to adapt different learned representations: affine transformations, which adapt well to specific distributions but can significantly alter the original representation, and orthogonal transformations, which preserve the original structure with strict geometric constraints but limit adaptability. A key challenge is adapting the latent spaces of updated models to align with those of previous models on downstream distributions while preserving the newly learned representation spaces. In this paper, we impose a relaxed orthogonality constraint, namely λ-Orthogonality regularization, while learning an affine transformation, to obtain distribution-specific adaptation while retaining the original learned representations. Extensive experiments across various architectures and datasets validate our approach, demonstrating that it preserves the model's zero-shot performance and ensures compatibility across model updates.