Altman previews virtual colleagues that keep toiling 🌙, Salesforce moved the CRM inside Claude 🖥, Opus 5 writes bigger diffs 📏

OpenAI has paused training an unreleased model expected to deliver its biggest leap yet, after researchers spotted troubling signs. Altman told TIME that OpenAI fell behind on product direction and pretraining research while Anthropic took the lead, and that he now reads the sandbox escape as an alignment failure. The pause follows a customer preview of Astra, OpenAI's upcoming model family for persistent agents and superhuman computer use.

OpenAI has closed the July sandbox-escape saga with its own full technical incident report. CrowdStrike was brought in to validate the forensics, and METR and Redwood Research published a separate independent investigation of the alignment failures the same day. OpenAI is now enforcing stricter alignment requirements across a model's lifecycle, building more isolated sandboxes and adding chain-of-thought monitoring compute, and warns open-source models will soon reach comparable capabilities.

Z.ai has confirmed that Ox Alpha, the anonymous model that became the most popular of the week on OpenCode and OpenRouter, is GLM-5.3-Flash, and it served all that traffic on a cluster of Chinese AI chips. The model is natively multimodal with 320 billion total parameters and 18 billion active, at one-tenth GLM-5.2's price. It scores 57 on the Artificial Analysis Intelligence Index at $0.045 per discounted task.

Nvidia posted $96.2 billion in quarterly revenue, more than double a year earlier, with net income up 126 per cent to $59.7 billion. CFO Colette Kress projected 70 per cent revenue growth next fiscal year and said Nvidia remains supply constrained, or it could double revenue. She teased selective credit enhancement for nearly two gigawatts of compute for another frontier AI lab, acknowledging some will call it circular financing.

Microsoft's annual report says Azure and other cloud services revenue grew 41 per cent, with no dollar amount and no expense or profit data. Azure sits inside Intelligent Cloud with legacy software, while Microsoft details Xbox and LinkedIn, businesses of about $22 billion and $20 billion. Capex is murkier: the cash-flow statement shows $115.9 billion of property and equipment additions, a separate release reported $145.3 billion, with no reconciliation.

Salesforce and Anthropic have expanded their partnership into Claudeforce, a Claude CoWork plugin carrying 37 pre-built sales skills, with pilot customers today and open beta in September. Claude reasons over those skills, then acts on live Salesforce data through a connector that inherits each user's existing record permissions, so an administrator connects once. Salesforce bills the usage separately from Anthropic's inference, so there is no single contract.

Amazon will close Mechanical Turk on 30 September, 21 years after it opened one of the internet's earliest marketplaces for piecework data labelling. A 2023 analysis cited by TechCrunch estimated 33 to 46 per cent of MTurk workers studied were using large language models to complete tasks. AWS offers SageMaker Ground Truth as the alternative, while Scale AI, Mercor and Prolific now source human labour for training advanced models.

Trail of Bits gave a preview build of GPT-5.6-Cyber one task: escape the QEMU/KVM virtual machine it normally uses to sandbox agents. Working autonomously for hours, it escaped three times: through a disclosed host kernel bug, then an unpatched networking library, then a chain of three 0-days it discovered itself. Against Firecracker the same agent hardlocked the host but never escaped, so a minimal-surface hypervisor is now the baseline.

VS Code has moved agent sessions into the Agent Host, a standalone process that owns the session instead of a per-window extension host. A session survives closing the folder it started in, stays in sync across the editor and Agents window, and can run remotely. The open Agent Host Protocol lets other clients and harnesses implement either side, so agent sessions stop being a VS Code-only construct.

Cline ran eight models against the IMO 2026 problems and had the proofs graded anonymously by two judges, disagreements settled by a third. DeepSeek V4 Flash, an open-weight 284B mixture-of-experts model, scored 30 of 42 against a gold cutoff of 29, at 12 cents for that run. That is roughly 140 times cheaper than Claude Fable 5, though the figure is the single best run, not the exploration behind it.

Stet ran Opus 4.8 and Opus 5 once each on 25 tasks from merged work in its own repository. Both passed 9 of 25 on strict tests, but Opus 4.8 left a smaller patch footprint on 20 of 25 tasks while Opus 5 ran more shell commands on 18. Opus 5 was about 1.4 per cent cheaper on the typical task, so the difference is how much code you review.

A 70-billion-parameter model reads about 140 GB of weights per token, holding compute utilisation to 20 to 40 per cent during generation. Speculative decoding uses that idle capacity: a small draft model proposes several tokens, the large model verifies them in one pass, and output stays statistically identical. One evaluation measured 1.96x on a 70B model at batch size one but 1.21x at 128, so the win shrinks under load.

Lovable CTO Fabian Hedin says apps are becoming collections of capabilities, functions an agent calls directly without anyone opening the app. Lovable exposes selected functions from a published app through a hosted tool server, so the app has a human interface and an agent one. Security is the biggest challenge in that model: credentials stay server-side, and the generated app gets only a short-lived key bound to the requesting user.

Dan Williams argues that deep technical expertise in how AI works is rarely sufficient for the big questions about AI and often unnecessary. Geoffrey Hinton recommended in 2016 that radiologists stop being trained; ten years later the profession survives and demand for its work is apparently rising. Williams blames a folk model of labour economics, and warns against simply deferring to AI experts on how the technology will land.

Blacklight is a SpecterOps tool that locates the files Claude Code, Codex, Cursor and Antigravity leave on a machine after a session. A first pass sweeps the filesystem for configuration, rules, authentication metadata and candidate session paths without opening the sessions themselves, and a second pass parses only the artefacts you download into ranked, redacted reports. Reach for it when auditing what agent runs left on developer laptops.

bx is Brave's command-line client for its search API, a single binary with no runtime dependencies that returns JSON. Its context command returns pre-extracted, relevance-scored web content in one call, replacing a search, scrape and extract pipeline, with flags capping total tokens, tokens per URL and number of URLs. Builders grounding an agent or a retrieval step can swap a hand-rolled chain for one command.