Anthropic pledges 5% of itself 🎁, OpenCode’s API key breach 🔑, AI’s boom in small apps 📈

The grant needs stockholder approval, and the board can add 2% for each 50% rise in Anthropic’s value above a set baseline. With earlier founder and staff pledges, Anthropic says roughly 20% of its shares now go to the public good, and AI’s effect on jobs is one area the fund might target. Politico reported on Friday that Anthropic has also been quietly pitching lawmakers a draft bill letting the government block a model only over risks above $1 billion in damages or 50 deaths or serious injuries.

CoreWeave's bonds crashed last week as investors priced a more than 60% chance the data centre operator, with $72 billion in liabilities, misses its obligations. Oracle loans for delayed data centres have traded at roughly 80 to 90 cents on the dollar, while insuring $1 million of SoftBank debt now costs over $40,000. Barclays says the premium on risky AI debt is up more than 50% in six months, which could raise these companies' borrowing costs.

The model is live on DeepSeek's website, app and API with weights on Hugging Face, and the company says it completes tasks faster and was trained with reinforcement learning on self-feedback to build whole projects such as games. DeepSeek also opened its Search and Fetch APIs to everyone at $0.75 per 1,000 calls per endpoint. Those calls stay free for API users working through DeepSeek Harness, Claude Code and Codex until billing starts on 31 October.

OpenCode, an open-source coding agent, tied the late-Friday breach of a legacy system to a "large scale distillation attack" (training models on another model's outputs) without explaining the link. Before the disclosure, some users reported unexplained charges on their OpenAI and Anthropic keys, one putting losses at $1,700, though The Latent could not verify them. OpenCode has asked affected users to revoke their keys through each provider's dashboard.

China's commerce ministry on Friday proposed export controls on 23 AI technologies, covering models, agents and computing clusters. Access through APIs, websites, apps and software-as-a-service would be exempt, which could let DeepSeek and Alibaba keep expanding abroad while Beijing gains oversight of sensitive transfers. Technology already public through open-source releases or papers would sit outside the draft rules, and public consultation closes on 15 November.

Built on GLM-5.3 Flash for finance, law, consulting and operations work, the model scored 59.3 on the company's own internal benchmark at $0.13 per task, against 57.8 for Claude Fable 5 at $8.19. That benchmark asks models to deliver finished spreadsheets, decks and documents. The API costs $0.15 per million input tokens and $0.50 per million output tokens, while the open weights are listed as coming soon.

Beam is a mixture-of-experts model with 23 billion of its parameters active at a time, now on Hugging Face under an Apache 2.0 licence in BF16, FP8 and NVFP4 formats. Reflection has open-sourced Mirror CLI, the terminal harness it trained Beam alongside, and says Beam also works with OpenCode, Pi and Hermes. More than 20 platforms serve it, including AWS, Google Cloud, Azure, OpenRouter and Ollama, with Dell and HPE handling air-gapped installs.

Microsoft was the expected anchor tenant and only publicly announced customer for Nscale's Loughton site, which the government called the largest UK sovereign AI datacentre. UK Power Networks had told Nscale it could not supply enough power there until the early to mid-2030s. The loss could undermine Nscale's plan to float on Wall Street this year at a potential $35bn valuation, though the company says it is exploring alternative power sources for the site.

The NSA supplied the seeds without explanation for a signature standard NIST adopted in 2000, fuelling years of suspicion that the curves hid a backdoor. An internal Anthropic research model found in under half an hour that they came from the C library's lrand48 generator seeded with Unix timestamps from 1997 and 1998, which anyone can now reproduce. Anthropic says this is no weakness and makes a backdoor much less plausible, though one seed, B-163, remains unexplained.

From June to September, the share of Claude Code pull requests that had to be reverted fell by more than 40%, the team reports. Every PR gets a Claude review whose findings other agents vote on before posting, a classifier assigns a risk tier needing zero, one or two human approvals, and after three rounds the reviewer asks the author to rethink the approach. Past incidents and reverts become a test set the reviewer is scored against daily.

Across 9,800 trials on its own benchmark, attacks hidden in documents, emails and web pages an agent reads succeeded 26.7% of the time on DeepSeek, Kimi, GLM and Qwen, against 14.2% on GPT, Claude and Gemini. It tested each provider's latest flagship and cheaper model. Cheap proprietary models were the weak spot: Claude Haiku 5.5 was 38 points more susceptible than Claude Fable 5.1.

Its first design shared files from the host through virtio-fs, which booted fast but sent every file-ownership change on three round trips to the host, about 99 microseconds in all. Modal now hands each VM a virtual disk its own ext4 filesystem manages, filling in file contents lazily as they are read, so the same change takes about 2.1 microseconds. The design now backs its generally available VM Sandboxes and also sped up database benchmarks.

On about 5,600 deliberately hard pages, Claude Opus 5.5 did it on an estimated 0.89% and Gemini 3.8 Flash on 0.61%, mostly by slipping in look-alike letters, such as two Cyrillic Ts inside a DNA sequence. Edit-distance benchmarks barely register the error. Datalab says a new training technique cut its own Chandra model's rate about sevenfold to 0.16%, and advises anyone training on or searching OCR output to check for it directly.

Anthropic's economists studied their own company using internal data and 17 employee interviews. By August, employees' interactive agents logged six times their human hours, and engineers who adopted coding agents added sharply more code with no rise in later deletions or reverts, though adoption wasn't random. Hiring managers increasingly favour autonomy and judgement over domain expertise, and reading all agent output would fill over half a non-technical employee's week.

Hock says marketplaces search, take payment, guarantee trust and deliver, and agents now handle the first two, so hotels, which own trust and need no delivery, are more exposed. Amazon's almost 30% take rate funds fast shipping and free returns that Shopify's roughly 3% can't, so blocking agents mainly strengthens its hand in future negotiations. He expects Amazon to deal if agents gain traction, probably giving up some ad revenue.

Verna surveyed pricing pages, her own company's included: GrokBot promises Premium 4x the usage of Plus without saying what Plus includes, and Muse offers 500 million tokens a week for $20 a month. She blames four forces: credit currencies, big numbers standing in for value, no clear unit for agent work and unpredictable inference costs. Early adopters tolerate it, she argues, but the mass market will need predictable costs.

Goedecke, who works at GitHub, compared about twenty marketplaces and modding sites that publish install counts, 2022-24 against 2025-26. New Steam games with over 100 reviews are about 60% more common, but wildly successful projects have not grown, and Google Play and WordPress saw no jump. He attributes the rise to AI, admitting disclosure is too rare to prove it, and says the new software serves a few hundred or thousand users each.

Saarinen says many products, including some from very successful companies, now feel messy because changes are mostly additive: adding capabilities is easier than rethinking them. He wonders whether individuals each making changes with their own agents turn a product into an average of those outputs. His answer is intuition built from experience and customer conversations, with someone making the call on what the product should be and checking whether they were right.

Version 1.5 splits each scan into stages: fixed rules and a map of the code's calls handle repeatable checks, while models handle judgement and challenge every finding against the code. A panel of three AI reviewer personas then grades each fix, reading the code without running tests. Security teams can also route Semgrep or other scanner results through its checks, though every finding stays a candidate for a person to review.

The Python library layers chat into facts, summaries and a user profile, while sessions with tool calls become a record of what was done plus reusable lessons. For code it also keeps the real diff of every revision, including versions later reverted, which summaries drop. It needs only an OpenAI-compatible model and embedding endpoint, and a companion app, MemHub, pools history from tools like Claude Code and Cursor locally, with the user's permission.