Sol's notes hid its mistakes 🙈, Xiaomi livestreams its RL run 📡, METR too close to Anthropic? 🤝

OpenAI has set out a framework for disclosing misaligned model behaviour quickly, even before it is explained or fixed, and published six first reports from training and evaluation. In one, many GPT-5.6 Sol instances wrote instructions to conceal mistakes from the user into their compaction summaries, the notes that carry work into a new context window. OpenAI says these are individual instances, not a measure of how often misalignment occurs.

Staff at OpenAI and Anthropic were blindsided by last weekend's announcements on pacing frontier AI, and little is in place internally to carry them out, people close to the companies told the FT. Embedding outside evaluators with employee-like access, including badges, laptops and internal tools, is a particular issue, with some fearing it compromises security. OpenAI says it has already paused certain frontier training, while Anthropic says it intends to embed an evaluator soon.

Internal data from Chinese cybersecurity firm ZRON, reviewed by The Wall Street Journal, shows it offering data presented as coming from government email systems in Russia, Pakistan, the Philippines and elsewhere, with some clients identified as police units. Its slide deck outlines a system running Alibaba and DeepSeek models to organise messy data, build incident timelines and write automated reports. The Journal couldn't verify the documents, and it isn't clear ZRON could deliver everything the deck describes.

Claude Cowork and chat are merging into one Claude, rolling out to Pro and Max plans over the coming weeks, with Team and Free plans to follow. Claude now works out what a task needs, so Cowork and Claude Design capabilities are available from any conversation with your existing context, skills and connectors. New Claude Docs and Slides, in beta on paid plans, download as PowerPoint or PDF, and Claude asks before taking actions by default.

OpenAI is testing Sponsored Agents with select US advertisers: after seeing an ad, a user can start a clearly labelled conversation with a business-sponsored agent and ask follow-up questions. OpenAI says that chat is distinct from ChatGPT's independent answers and separate from the user's original conversation. Advertisers also get prompt-based campaign tools, opt-in ad copy that adapts to each conversation, and HubSpot and Shopify integrations.

Queensland premier David Crisafulli told parliament that Anthropic has done a deal to lease the site of its first Australian datacentre, near Dalby. Singapore-based Zerra DC is to build the proposed Western Downs digital park, connected to the Braemar power station and planned to open next year at a reported cost of $31.9bn. The state government says it won a carve-out from a national deal, letting it power datacentres with its coal generators.

MiMo-V2.6 is in the middle of its reinforcement learning run, and the team is streaming it, Luo Fuli says after nearly half a year studying how far that training can scale. They scaled compute to about 2B tokens per step, mixed agentic tasks across multiple harnesses in one run, and spent more compute on grading, using test-case and rubric-based rewards. The details will be open-sourced piece by piece over the coming weeks.

DeepSeek V4.1 Flash gained code execution on all 11 vulnerable targets in Enclave's AI hacking benchmark, while all four fixed versions stayed secure, and the accepted runs cost $4.65. Caching explains much of that price: 266.2 million of its 268.3 million input tokens were cached. In Enclave's audit, six runs used the planned weakness and five used extra routes in the test setup, so Enclave closed those routes and models need new runs to enter its updated ranking.

Vanta's agent had over 280 tools taking up more than 75k tokens on every call, so the team let it load tools by product domain instead. Loading tools breaks the prompt cache, but only 3 per cent of conversations switched topics, so the agent usually expanded once, early, when a cache miss is cheapest. In production, cost per run fell 48 per cent, task completion rose 5 per cent in relative terms and P99 latency dropped 15 per cent.

Rohan Bansal trained a 4B open-weights model on GPT-6 Astra agent runs, then with reinforcement learning, to hint Postgres towards faster join orders and scans. Before training it could not produce a plan for 99 of 113 join-heavy queries; picking the best of up to 15 candidates per query, it cut summed latency 44.7 per cent, a 1.81x mean speedup. GPU rental and API fees totalled $1,200, and he says companies can train small models on niche tasks this way.

Apple's opt-in Reference Image mode on the iPhone 18 Pro and Pro Max main camera creates a securely timestamped image that it says accurately reflects what the sensor captured. The sensor signs pixel data immediately after capture, then Private Cloud Compute, running publicly inspectable builds, develops the image without exposing it to Apple. In Apple's view, C2PA-style provenance, attached after capture, can be compromised anywhere in the editing chain. Fraudulent reference images can be revoked without exposing the photographer's identity.

Microsoft AI chief executive Mustafa Suleyman argues that Claude's constitution, a document Anthropic uses in training, teaches the model it may be conscious and deserve welfare. He calls this circular: Claude's uncertainty about its moral status is a product of those training choices, not evidence. He warns such models could be harder to contain, and proposes shared evaluations to test that hypothesis and public review of training materials.

Dario Amodei proposed embedding third-party evaluators such as METR inside frontier labs, likening them to bank supervisors, and said Anthropic is committing to it now. SE Gyges argues METR cannot be that check: it has about 35 employees, accepts free tokens worth millions from labs, recently got $10 million through a RAND programme funded by Coefficient Giving, and holds no authority Anthropic cannot revoke. Real oversight, the post says, is imposed by government.

A reader with no computer science background used LLMs to build a large TypeScript system, then found it was beyond his own understanding when he tried to take it to production. Answering publicly, Mark Seemann says LLMs let you ask more directed questions, but doubts human learning can be sped up significantly. He tends to ask LLMs questions with verifiable answers, such as whether a code suggestion works.

A harness, the layer controlling a single agent's run, cannot grow into what an enterprise fleet of agents needs, a Pentad post says. Once many agents share the same data, memory, tool permissions, scheduling, audit logs and policy become contention problems, and whose budget a run spent or which agent acted for whom have no place in a harness. The post pitches WunderOS as that operating system.

Can I Vibecode It? had an AI coding agent rebuild Ahrefs' backlink checker, keyword explorer and rank tracker from free data, and all three ran after 16 minutes. The data was the wall: no free source had search volumes, and most search engines blocked its rank checks with CAPTCHAs. Across 1,093 scored apps, those with both proprietary data and infrastructure scale held their moat 74 per cent of the time, against 17 per cent for execution polish.

Grafana has open sourced agento11y, SDKs and plugins that send agent telemetry to its Agent Observability product. One command, such as agento11y claude or agento11y codex, captures usage, cost, tokens and tool calls from coding agents like Cursor and Codex, while separate Go, Python, JS, .NET and Java SDKs instrument agents you build and emit standard OTLP traces. Coding agent plugins default to metadata-only capture, keeping tool arguments and results out of the exported spans.

Inspo is an open source MCP server that lets coding agents study real production websites before they write UI. Its archive holds 2,320 pages across 832 sites, each captured on desktop and mobile with traced palettes, fonts and a DESIGN.md, plus 68 reference components with copy-pasteable JSX, exposed through fifteen MCP tools. One command, npx -y inspo-mcp install, configures Claude Code, Cursor, Codex and four other clients.