- Funding & deals
The New York Times' Hard Fork looks at what Anthropic's S-1 means for the AI boom. The episode calls Monday's S-1 the clearest view yet into the business side of the AI boom, ahead of what could be the largest IPO in history.14 It also discusses the midterms and Microsoft's buyout offer to 60,000 employees.14
- Industry
Wired reports that a defense contractor connected Claude Code to a secure Pentagon system. On its Uncanny Valley podcast, Wired discussed its scoop on a contractor who gave a personal Claude agent access to a secure Department of Defense system, calling it one of the most outlandish government uses of AI the team has seen.15 The episode also went through Anthropic's S-1, which shows first-half 2026 revenue of $16.3 billion.15
- Models
A new Claude Code rule now blocks recreations of specific existing games. After users reported reverse-engineering refusals last week, a rule sent from Anthropic's servers around October 27 under the flag tengu_heron_brook tells Claude not to build a recreation of a specific existing game, even if the result is renamed.16 Original games, genre games and games built from the user's own designs are still allowed, though Reddit users say new safety classifiers are also blocking other requests.16
- Products
ChatGPT shopping results now include a virtual try-on button. Since October 1, users can tap Try on beside clothing in ChatGPT shopping results and upload a photo to see a generated preview of themselves wearing the item, and the photo stays saved until they delete it.17 OpenAI says the previews do not guarantee fit or size.18
- Policy & legal
A coalition backed by Cate Blanchett tells the court that licensing for AI training is workable. In a 31-page amicus brief, the Human Consent Foundation, CAA and other creative groups dispute claims by OpenAI and Microsoft that licensing works for training is impossible, pointing to collective licensing groups such as BMI.4 OpenAI pointed instead to a Friday blog post listing briefs that support fair use.19
- Models
Doximity says widely used models get more than 2 percent of drug dosing claims wrong. Doximity tested nine AI models on more than 8,000 dosing questions and found at least 2.3 percent of dosing claims were wrong, with 43.8 percent of detected errors rated serious or potentially lethal, and it released the results alongside its own DoseCheck verification tool.20, 21 Models struggled most with recently approved drugs and with adjusting doses for age or kidney function.21
- Models
The Biological Computing Co. released Mischief 1, an image-to-video model built on Wan 2.2. TBC says techniques learned from lab-grown neurons make the model about five times faster than Wan 2.2 at the same quality, though it has published no independent benchmarks.22 It costs $0.005 per second of video on fal and OpenRouter.22
- Policy & legal
A Forbes column says Beijing plans export controls on AI that mirror US rules. Citing Reuters reporting, the column argues that China's plan to copy the US export controls imposed since 2022 shows those controls backfiring on the United States.23
- Dev tools
Security agents from Anthropic, OpenAI, Cloudflare and Google all use the same find-then-disprove design. MarkTechPost's comparison finds that in each tool, one agent reports a bug, a separate agent tries to disprove it, and only findings that survive get a patch.24 The tools differ in access: Cloudflare's skill is MIT-licensed and model-agnostic, while OpenAI's Codex Security CLI is Apache-2.0 but still needs OpenAI scanner access.24
- Models
Samaya AI's Sterling, a model with 3B active parameters, beats Claude Fable 5.1 on its own finance benchmark. Sterling is a LoRA post-train of Nemotron 3.5 Lightning and scores 58.9 percent on FrontierFinance, against 55.9 percent for Fable 5.1, at $0.61 per question instead of $6.89.25 All results are self-reported on Samaya's own benchmark.25
- Funding & deals
Databricks and SAP are bringing four more Databricks tools to SAP Databricks. Genie, Lakeflow, Unity Gateway and Lakebase are coming to SAP Databricks, and Genie will link with SAP's Joule assistant, though neither company gave dates.26 SAP also introduced Application Data Foundation to make data in its business software easier to reach.26
- Research
Veracode reports its first real gain in the security of AI-generated code. Across 18 models, the average security pass rate rose from 56 to 62 percent, and Claude Opus 5.5 and GPT-6-Sol tied for the lead at 73 percent.27 Most of the gain came from cross-site scripting, while flaws that need data-flow reasoning, such as log injection, still trip up the models.27
- Industry
A hijacked Claude Code extension on Cursor's marketplace tried to give attackers persistent access. Fast Company reports that the malicious extension tried to give hackers lasting access to developers' computers, adding to concerns about the security of AI coding tools.28
- Models
Edge0 released Audio8-TTS-Realistic, an open-weight voice cloning model with 0.6B parameters. The Apache 2.0 model supports 23 languages and reports a 4.86 percent average error rate on CV3, ahead of Qwen3-TTS-1.7B.29 It is still weak in lower-resource languages, with Thai near 49 percent word error rate.29
- Products
Canva announced its own vision-language model for design. Fast Company says the model is Canva's attempt to move AI-generated graphics past the cluttered, cliché-filled style that marks so many AI posters.30
- Research
Researchers from Meta, CMU and UW filtered training data to cut jailbreak success by up to 22 percent. Their SimpleLoRIF method ranks training examples by how closely their gradients align with unsafe outputs, and it produced SafeTulu, a filtered version of the Tulu 3 mixture.31 Models from the Qwen, Llama and OLMo families trained on SafeTulu resisted GCG attacks better while matching or improving math scores.31
