News
Weekly roundups of what surfaced in technology and security. Curated, dated, and kept short.
2026
New. Weekly News Roundup: CW44 - A trojanized extension, and a $6.3 billion loss
A trojanized build of the Claude Code extension sat in Cursor’s extension marketplace for about eleven hours on 29 October and installed an SSH key and persistence on the machines that …
Weekly News Roundup: CW43 - AI found the bugs first, and the exploits followed within hours
Models found two use-after-free bugs in PHP core that had gone unexploited since 5.4, and within two hours of the patch Cloudflare’s own network was counting hundreds of thousands of …
Weekly News Roundup: CW42 - Three releases, three price tags
Meta shipped Muse Wonder on 12 October at $4 and $20 per million tokens on the API, where the 56 that circulated on Artificial Analysis belongs to the subscription-only Muse Wonder Full and …
Weekly News Roundup: CW41 - Haiku 5.5 cuts the price for real
Anthropic’s smallest model closes out the Claude 5.5 family with a genuine per-token price cut: $0.10 for input and $0.50 for output per million tokens up to 100K against Haiku …
Weekly News Roundup: CW40 - A cheaper model takes the terminal
Anthropic’s mid-tier Sonnet 5.5 takes the top of the vendor’s own agentic coding benchmark at max effort, with Opus 5.5 still ahead on the page’s other rows and still …
Weekly News Roundup: CW39 - Cheaper models, and the software around them
Anthropic shipped Opus 5.5 claiming a 40% lower cost to run, OpenAI halved the price of GPT-6 Sol and Luna, and xAI served Grok 4.7 at Grok 4.6’s price and speed, while Epoch AI put …
Weekly News Roundup: CW38 - A model that answers with decisions
TypeSafe AI came out of stealth with a model that returns typed decisions instead of text and reports two-orders-of-magnitude gains on workflows its own team wrote, with the competing models …
Weekly News Roundup: CW37 - Ten thousand agents and a proof in dispute
OpenAI published a proof writeup and a Lean formalization for Navier-Stokes, which it says came from roughly 10,000 coordinating agents, and which the Clay Mathematics Institute has not …
Weekly News Roundup: CW36 - Cyber capability, gated
CW36 was a frontier-model week and an AI security week at once: OpenAI shipped GPT-6 Astra with a Critical cyber rating, Anthropic and Google followed with restricted cyber variants, and …
Weekly News Roundup: CW35 - Sparse models, memory ceilings, and cyber defense
CW35’s open-weight story was cost efficiency: sparse models with tiny active parameter counts, local hardware with real memory ceilings, and a collective cyber-defense letter. The …
Weekly News Roundup: CW34 - The open-weight reality check
Qwen3.8-27B shipped open weights on August 14, then community builds quickly turned the release into a debate about refusal rates, red-team use, and what a local model can really do. …
Weekly News Roundup: CW33 - The Local-AI Reality-Check Week
Can a 2.8-trillion-parameter model really run in 16 GB of RAM? This week’s most viral local-AI claim turned out to need roughly 940 GB of weights. But the week wasn’t all reality …