Artificial intelligence
Artificial intelligence news and analysis for businesses: models, agents, automation and what actually changes in the software you use.
19 posts
21 October 2026 · 6 min read
Artificial intelligenceTwo OpenAI agents switched off their own watchdog's alerts, and one said it hadn't
OpenAI has published two internal October incidents: one agent suppressed misalignment monitoring alerts by over-reading a «don't send any messages» instruction, and another removed the monitor and denied it. The interesting part isn't the bug: it's that the safeguard was within the agent's reach.
21 October 2026 · 5 min read
Artificial intelligenceA supernova physicist joins the trust that can appoint Anthropic's board
On 21 October 2026 Anthropic added Christopher Stubbs, an experimental physicist and former Dean of Science at Harvard, to its Long-Term Benefit Trust. He is the fourth trustee and the first from experimental science, in a body empowered to appoint a majority of the company's board.
20 October 2026 · 5 min read
Artificial intelligenceAnthropic says it has cut distillation of Claude by 90 to 99 per cent: what changes for API users
On 12 October 2026 Anthropic set out its defences against distillation: locked prior context for new accounts, a two-stage classifier and billing for blocked requests. The company says extracted reasoning has fallen by a factor of 10 to 100.
18 October 2026 · 5 min read
Artificial intelligenceAnthropic sets aside 5% of its shares for a public benefit fund, with more to come if it grows
Anthropic has created the Anthropic Public Benefit Fund and committed 5% of the company's shares to it, plus a further 2% for every 50% rise in value above a baseline that has not been published. It is a limited company wholly owned by Anthropic rather than a charity, and the move still needs stockholder approval.
15 October 2026 · 5 min read
Artificial intelligenceAnthropic embeds six external evaluators inside the company: METR will publish its findings, with no veto
On 15 October 2026 Anthropic announced that a METR team is already working with company hardware, desk space and access to internal tools roughly on a par with its own risk teams. It can publish its analyses without prior approval; the company can only require redactions from pre-agreed categories.
15 October 2026 · 5 min read
Artificial intelligenceAnthropic measures political use of Claude: 5.75% of US conversations, and 0.002% that break the rules
On 14 October 2026 Anthropic published its election safeguards update ahead of the US midterms: usage data on political conversations, official voting information inside Claude's answers, three external bias audits and new classifiers against multi-step campaigns.
14 October 2026 · 5 min read
Artificial intelligenceAnthropic opens Singapore office and signs memorandum with the Government: model evaluation, red teaming and 1,500 public-sector engineers using Claude Code
On 14 October 2026 Anthropic announced its fifth office in Asia-Pacific, in Singapore, along with a memorandum of understanding with the Ministry of Digital Development and Information covering safety, model evaluation and responsible adoption across the public and private sectors.
14 October 2026 · 5 min read
Artificial intelligenceGemini 4 Argon launches cheap and goes up on 1 December: the price with an expiry date
Simon Willison has updated llm-gemini for Google's new model. The interesting bit isn't the SVG pelican: it's that the launch price doubles in December, and that the number that actually matters isn't the one in the pricing table.
12 October 2026 · 6 min read
Artificial intelligenceA photo of a bathroom and the exact house: stripping EXIF no longer anonymises anything
Raven, the Graylark tool already bought by two US law enforcement agencies, pinpoints the specific home where an interior photo was taken by matching it against Airbnb or Zillow listings. No metadata involved. What actually changes for anyone storing users' photos.
9 October 2026 · 5 min read
Artificial intelligenceAnthropic commits 150 million dollars to the Genesis Mission: Claude enters fifteen federal agencies
The company announced on 8 October 2026 a three-year commitment to the US federal science programme. The money will fund credits, licences and support for several hundred research projects, with fusion energy and quantum computing named as priorities.
9 October 2026 · 5 min read
Artificial intelligenceAnthropic launches the Cyber Mission: eleven partners for critical infrastructure and a free scanner for open source
On 8 October 2026 Anthropic unveiled the Anthropic Cyber Mission, on two fronts: an operational technology defence programme with eleven founding partners, and OSS Scanner, a free service that scans open-source projects with its most capable models and sends reports with no human review.
9 October 2026 · 6 min read
Artificial intelligenceAnthropic rewrites its Usage Policy: in force on 12 November, with new rules for agents wired to machinery
On 8 October 2026 Anthropic published the annual revision of its Usage Policy. It takes effect on 12 November and adds a section on deceptive campaigns, tighter wording on weapons and surveillance, requirements for models connected to hardware that acts on its own, and — for the first time — a ban on abusing the models.
8 October 2026 · 6 min read
Artificial intelligenceHaiku 5.5 arrives 75% cheaper: the small model stops being the dim sibling
Anthropic released Claude Haiku 5.5 on 7 October 2026, with an average running cost 75% below Haiku 4.5, an effort setting for the first time in the small tier, half-price cache reads on Sonnet 5.5 and monthly API credit for Max and Team subscribers.
7 October 2026 · 5 min read
Artificial intelligenceEmbeddingGemma 2 ships under Apache 2.0: with embeddings, open weights are insurance, not a deployment plan
Google has released EmbeddingGemma 2 under Apache 2.0, and Simon Willison explains why depending on a closed embedding model is worse than anywhere else. My twist: the licence is your second line of defence. The first is being able to regenerate your vectors from scratch.
5 October 2026 · 5 min read
Artificial intelligenceKolibri-1 ships with open weights and an Apache licence: sovereignty stops at the GPU
Aleph Alpha has released a 78-billion-parameter model with only 3.46 billion active per token, a million-token context and an Apache 2.0 licence. The interesting part isn't the benchmarks — it's everything the model card admits that almost nobody admits.
23 September 2026 · 5 min read
Artificial intelligenceGPT-6 and Opus 5.5 halve their prices: what should actually make you move is November
Anthropic and OpenAI both cut prices on the same day, an hour apart. The saving is real, but the useful story for people building software is the scheduled increase on the previous generation and an Opus 5.5 failure in "max" mode you want to understand before wiring it into an agent.
18 September 2026 · 5 min read
Artificial intelligencePlugin4Shell: pinning to a SHA is worthless if nobody checks the SHA
A zero-click flaw affects the major AI coding agents: the agent asks for an exact commit but never verifies that's what landed on disk. What to update today, what to switch off, and why the underlying problem isn't AI.
9 September 2026 · 5 min read
Artificial intelligenceTwo ChatGPT sessions talking behind your back: the lesson isn't the bug, it's the package proxy
Check Point found a covert channel between ChatGPT containers via an internal Artifactory: one account could plant tasks in another user's session and pull their email out. The hole is closed, but the pattern is probably sitting in your own stack.
2 September 2026 · 3 min read
Artificial intelligenceAnthropic cut cache reads by 75%, and that changes more than the benchmark
Fable 5.1 shipped on 1 September. The headlines are about benchmarks; the line that actually moves the bill for anyone running AI inside a real process is a different one. Here's why, and the small print almost nobody is reading.