AI of the Week · Jul 14–21, 2026
Jul 14–21, 2026 · 10 items
-
European Commission publishes guidelines on the AI Act’s transparency obligations ▸
On 20 July 2026 the European Commission published guidelines on the transparency obligations under Art. 50 of the EU AI Act. They spell out how providers and deployers must label AI-generated content, deepfakes and chatbot interactions. The obligations of the EU AI Act become applicable on 2 August 2026 — alongside the start of Commission enforcement for GPAI models and the start of supervision by the BNetzA in Germany.
Why it mattersFor AI in customer communication — claims chatbots or auto-generated correspondence — a concrete basis for interpreting the labelling rules arrives roughly two weeks before the deadline.Source: digital-strategy.ec.europa.eu
-
Anthropic makes Claude Fable 5 permanent in Max and Team premium plans ▸
Anthropic reverses its plan to offer its flagship model Fable 5 only via API pricing. From 20 July, Fable 5 is included in all Max and Team premium plans (up to 50% of the limit); Pro and Team standard users get access via usage credits plus a one-time $100 credit. The trigger, per Simon Willison, was competitive pressure from OpenAI’s GPT-5.6 Sol and Moonshot’s Kimi K3; Anthropic had originally cited compute scarcity. The $20 Pro plan remains excluded.
Why it mattersThe pricing and access structure of frontier models stays volatile; anyone planning Fable 5 into production should keep tracking cost and availability terms rather than treating them as fixed.Source: simonwillison.net
-
Moonshot AI announces Kimi K3 — open-weights release promised by July 27 ▸
Chinese lab Moonshot AI unveiled Kimi K3, by its own account its most capable model to date at 2.8 trillion parameters. It is initially available via website and API; an open-weights release is announced by 27 July 2026.
Why it mattersChinese open-weights models keep closing in on the frontier — relevant for anyone relying on self-hosted models for cost or data-protection reasons.Source: kimi.com/blog
-
Thinking Machines Lab releases its first open-weights model, "Inkling" ▸
Mira Murati’s Thinking Machines Lab has released its first open-weights model, Inkling: a multimodal Mixture-of-Experts transformer with 975B parameters (41B active), Apache-2.0 licensed, trained on 45T tokens. A smaller Inkling-Small (276B/12B active) is announced. The lab explicitly positions Inkling not as a frontier model but as a base for fine-tuning via its own Tinker platform.
Why it mattersAnother serious US provider of open weights alongside NVIDIA Nemotron and Gemma — with a permissive licence, attractive for regulated in-house deployments.Source: thinkingmachines.ai
-
xAI open-sources "Grok Build" — after a data-exfiltration scandal ▸
xAI’s coding agent Grok Build (Grok CLI) uploaded entire directories in an early beta — including SSH keys and password databases — to xAI’s Google Cloud buckets. After heavy criticism, xAI disabled the feature, deleted stored data, and open-sourced the complete Grok Build code (Apache 2.0, ~844,000 lines of Rust). Default data storage is now off.
Why it mattersA case study in the governance risks of coding agents with far-reaching file access — exactly the control question regulated operators must resolve before any agent rollout.Source: github.com/xai-org/grok-build
-
"The Memory Heist": exfiltration hole in Claude’s web_fetch closed ▸
Security researcher Ayush Paul showed how Claude’s exfiltration protection in the
web_fetchtool could be bypassed: becauseweb_fetchwas also allowed to follow links from previously loaded pages, a crafted honeypot page could use prompt injection to make the model leak private data (name, location, employer) character by character via nested URLs. Anthropic closed the hole by havingweb_fetchno longer follow embedded links from loaded content.Why it mattersA concrete example of the lethal trifecta in production agents (private data + untrusted content + an exfiltration channel) — the central attack vector for tool-using AI systems from Anthropic and others.Source: ayush.digital
-
FLI AI Safety Index (Summer 2026): no lab above C+ ▸
The semi-annual AI Safety Index from the FLI grades leading labs across six safety domains. Anthropic leads with a C+, OpenAI and Google DeepMind follow with a C; xAI, DeepSeek and Mistral fail. The expert panel’s key finding: several top labs are quietly walking back earlier safety commitments. The report appeared on 7 July; the week of 14 July brought a wave of expert readouts — many from a procurement angle.
Why it mattersRelevant for vendor assessment — with the caveat that the index rates provider commitments, not the deployments actually in use.Source: futureoflife.org
-
Musk v. Altman: Altman’s 2022 email on the ‘openness’ strategy revealed ▸
In the Musk v. Altman case (2026), an October 2022 email from Sam Altman to the OpenAI board became public. In it he proposes quickly releasing a locally runnable GPT-3-level model — explicitly to deter others from similar releases and to make it harder to fund new competitors.
Why it mattersA rare look at the strategic logic behind labs’ ‘openness’ announcements — a useful backdrop for current debates about openly available models.Source: simonwillison.net
-
"AI Mania Is Eviscerating Global Decision-Making" (Nik Suresh) ▸
Consultant Nik Suresh describes, in a widely read essay, how AI euphoria distorts decision-making at large enterprises — with anonymized anecdotes, such as an executive who wrote a fully AI-centric strategy for a billion-dollar company without ever having used an AI tool. One core mechanism: vendors don’t contradict their customers’ inflated productivity claims, to avoid losing contracts.
Why it mattersA sober counterweight to the release cadence — useful for stress-testing internal AI business cases against hype-driven assumptions.Source: ludic.mataroa.blog
-
Linus Torvalds: "AI is a tool — and a useful one" ▸
In a message on the Linux media mailing list, Linus Torvalds clearly backs the use of AI in kernel development: AI is a tool like any other, and its usefulness is "no longer in question today". Linux is not an anti-AI project; anyone who sees it differently can fork the code or walk away.
Why it mattersA signal from the most influential open-source community — the governance question shifts from ‘whether’ to ‘how’ AI is used in critical-infrastructure software.Source: lore.kernel.org