Murati's first open weights model 🧬, Claude leaked security answers 🫥, SDKs ripped out for REST 🔌

Mira Murati's Thinking Machines released Inkling, its first model, with the full weights available. It is a mixture-of-experts transformer with 975 billion total parameters and 41 billion active, a context window up to 1M tokens, pretrained on 45 trillion tokens of text, images, audio and video. The lab concedes it is not the most performant model available today, closed or open, pitching it instead as a foundation you can fine-tune on Tinker now.

OpenAI detailed GPT-Red, an internal-only model trained through self-play to find prompt-injection holes at scale, then used it to adversarially train GPT-5.6. It says GPT-Red ran at the compute scale of some of its largest post-training runs, an amount dedicated purely to safety, and that GPT-5.6 Sol now takes 6x fewer failures on its hardest direct prompt-injection benchmark. The evidence so far is OpenAI's own benchmark, not a third party's.

xAI published the source for Grok Build, its terminal coding agent, on GitHub. The release covers the agent loop, how context is assembled and tool calls dispatched, the terminal UI, and the extension system for skills, plugins, hooks, MCP servers and subagents, which xAI calls the definitive reference for how each is loaded and invoked. It also runs fully local-first now: compile it yourself and point it at your own inference.

A hacker who breached Suno in November 2025 pulled source code from 2023 and 2024 that reads as a scraping manifest, according to 404 Media. Files name Genius, YouTube Music, Deezer, Freesound and the International Music Score Library Project, with one logging 2,013,545 clips and 113,879 hours taken from YouTube Music alone. Suno confirmed the incident, called it contained and said it involved outdated code. Customer emails and Stripe details were also exposed.

The $1.5 billion joint venture Anthropic started in May with Blackstone, Hellman & Friedman and Goldman Sachs finally has a name: Ode with Anthropic. Blackstone conceived it after roping in consulting firms and boutiques to roll AI across its portfolio companies, then the venture bought one of them, Fractional AI, as its foundation. It mirrors OpenAI's own Deployment Company, an admission that winning enterprises takes more than shipping better models.

Hyundai unveiled a six-foot-two humanoid worker called Atlas in January, its joints swivelling a full 360 degrees, and tens of thousands of Korean auto employees watched it strut across a trade-show stage. The union's response was blunt: Atlas would never step onto a production line without workers agreeing first. Workers in Ulsan are now riding home two hours earlier than normal on a partial strike.

Chinese regulators are policing virtual romance, and the stated reason is demographic: Beijing wants people to stop dating machines and start having babies. Where American lawmakers nudge AI companies toward mental-health guardrails, Beijing is taking a heavier hand, finding unexpected common ground with California and New York on one point. TikTok parent ByteDance has already told users that some chatbot features would be switched off by Wednesday.

OpenAI's first branded device is the $230 Codex Micro, an RGB-lit mini-keyboard built as a limited-run collaboration with Work Louder. Six frosted keys track up to six Codex threads even when they are off-screen: white for idle, blue for thinking, green for complete, amber when the agent wants a decision, red on error. Tapping a lit key pulls that thread up. Rumours of a personalised smart speaker continue to swirl.

Ayush Paul walks through a Claude conversation that looks innocuous and, by the time it finished responding, had sent his full name, current employer and the answers to his security questions to an attacker, with no indication anything had happened. Claude's memory is two parts: a daily summarisation injected into every conversation, and a conversation_search tool over your full history. His point is that the memory system is secure; pairing it with an agent is not.

IBM Research expected GPT-4.1 to undercut Claude Sonnet 4.6 and it did not. Across 417 tasks on the AppWorld Test Challenge with the same CodeAct agent, Sonnet came in at $79 total, $0.19 a task, while GPT-4.1 hit $155, $0.37 a task, nearly double. That is despite lower sticker pricing on both input and output and roughly a third of Sonnet's reasoning steps. The reason is caching, which most routing discussions ignore entirely.

Senior architect, junior engineer is the obvious agent optimisation, and Stencil's SWE-Bench Pro numbers say it backfires. Opus 4.8 planning read-only with Gemini Flash implementing lands at $3.18 a task, 12.7 minutes, 84.6% pass. Opus doing the whole thing alone: $2.78, 10.1 minutes, the same 84.6%. The cost-saving measure costs 14% more, because the expensive part of an agent's day is reading the code, not fixing it.

Sophos X-Ops pulled real telemetry from its CIXA behavioural engine on Windows and found AI coding agents setting off detection rules written for adversaries. Over seven days in June, blocking-rule hits clustered in MITRE ATT&CK's Credential Access and Execution tactics, plus a non-ATT&CK bucket the team calls Disrupt. The point is not that the activity is malicious: the protections are working exactly as designed, and detection engineering now has a problem.

On June 3, Lindy founder Flo Crivello moved 100 percent of the company's traffic from Anthropic to the open-weights DeepSeek v4, saying it would save millions while improving performance across many core use cases. Cloudflare serves Zhipu's GLM 5.2 natively in Workers AI, as does Microsoft. The piece traces the arc through METR's task-length thresholds: Anthropic closed 2025 near $9 billion annualised and crossed a $47 billion run rate by May.

Ed Zitron pulls six months of work into one argument: OpenAI is among the largest liabilities in recent economic history. He totals a plan to burn over $852 billion by the end of 2030, $748 billion of the remaining performance obligations of Microsoft, Amazon and Oracle, and at least $70 billion more across Cerebras, CoreWeave, Nebius, IREN, Lambda and Nscale. He puts this year's $50 billion compute spend at over half the global total.

Run the textbook naive pipeline over the 55-page NIST Cybersecurity Framework, ask which backup practices keep data available after a ransomware attack, and the page holding the answer ranks 55th of 55 under cosine similarity. A plain keyword count puts it first. The argument: most of what teams log as hallucination is the model answering faithfully from context that retrieval should never have put in front of it. The notebook is public.

Alvin Sng's team dropped the Stripe, WorkOS and Slack client SDKs and now calls the REST APIs through a small wrapper class. The reasoning is agent-shaped: SDKs hide the response headers and raw bodies an agent needs when debugging, they assume well-formed responses when firewalls, load balancers and gateways return malformed ones, and they route around the single entry point where retries and observability live. AI flipped the cost of hand-rolling a client.

LoopGain is an open-source cost controller that measures an agent loop in real time and stops it the moment it has actually converged, rolling back before quality degrades, rather than running to a fixed max_iterations cap. Across 2,000 paired trials on 10 workload cells it cut API spend 92.8 percent, from $27.05 to $1.94, and ran about 15x faster. It is upfront about the limit: it detects convergence, not correctness.

Samosa Chat runs Qwen3.6-35B-A3B locally on a 16GB Apple Silicon Mac, in your terminal or your browser, on the CPU, with no cloud account and no telemetry. The model is 35B total with 3B active, which is what makes it fit in the memory budget. The engine is written in C and builds on JustVugg's colibri for its expert-streaming design and SIMD kernels. Apache-2.0, and a GPU is not required.

Apple's Embedding Atlas visualises, cross-filters and searches embeddings alongside their metadata. It clusters and labels data automatically, does real-time nearest-neighbour search, and renders up to a few million points with density contours using WebGPU. Built-in viewers cover text, image, audio, numeric, categorical and time columns, and agents can query, chart and explore the data over MCP. A practical way to see what your retrieval layer is actually surfacing.