Researchers compiled the libsignal reference implementation with the standard Rust compiler for AArch64 Linux. A human-approved specification states what the compiled program must guarantee; AI agents produce proofs that Lean checks.
The technical paper reports about 6.5 million lines of checked proof code. About 26,000 lines of specifications and machine semantics remain trusted inputs.
Its proofs address the post-quantum handshake and message ratchet, including message confidentiality, authentication and recovery after key compromise. The authors reintroduced two fixed bugs and showed that their symbolic security claims fail for the altered binaries.
The formal artifacts are public. The team says the work took four weeks, with people selecting and reviewing what the agents had to prove. An agent is a model running a multi-step task with tools.
Caveat: The guarantees depend on human-approved specifications, modeled hardware behavior and assumptions about cryptographic routines outside the proof; they do not certify every Signal build or feature.
(HN 532 points · 169 comments · Google DeepMind paper and proof artifacts)
Google says Carbon is available through Google AI Studio’s API and Gemini for business for paying accounts. It is also rolling out to AI Ultra subscribers in the Gemini app and Antigravity.
Introductory API prices are $2 per million input tokens and $10 per million output tokens, units of information processed and generated. Google’s footnote says these rise to $4 and $20 after November 30.
Antigravity’s /swarm preview assigns parallel workers separate git worktrees and includes agents that challenge changes by running tests. Google warns that parallel teams use quota faster than one chat.
Caveat: Google’s performance and cost comparisons use selected tests, so measure your workload and the full quota consumed by parallel workers.
(Techmeme · Google availability announcement and Antigravity documentation)
Scale Labs and the Center for AI Safety report a 41.30% acceptance rate for Fable 5.2, versus 21.25% for Opus 5.5 and 17.92% for Fable 5.1. A benchmark is a fixed set of tasks for comparing models.
The test has 240 sourced freelance projects; its leaderboard scores 230 private projects. Reviewers inspect delivered files against human work, and the reported score uses the best of two available operating environments.
The test covers deliverables including audio, 3D work, documents and code. The test description says agents also have access to specialized media tools and a $30 generation budget per task.
Yesterday Afternoon’s release item covers Fable’s access, pricing and separate coding tests. This independently reported result is the new development.
Caveat: Acceptance by the test’s reviewers does not mean a model can replace a freelancer across projects involving clients, physical work or long-term support.
(Techmeme · Scale Labs leaderboard and collected analysis)
A tool for modern web code from coding assistants#
The project’s CLI installs skills covering modern browser APIs, accessibility, performance and compatibility. The documented search, retrieve and legacy-scan commands can surface a specific guide or flag older patterns.
The Chrome documentation gives examples for native dialogs, container queries, passkeys and page-loading performance. The repository is maintained with contributors from Chrome, Edge, Firefox and MDN.
The repository predates today’s HN discussion; its new visibility comes from the HN submission rather than a launch this afternoon.
Caveat: The CLI reports anonymous install and guide-use telemetry by default; the README documents the DISABLE_TELEMETRY=1 opt-out, and generated code still needs browser testing.
(HN 167 points · 46 comments · project README and Chrome documentation)
Mia Sato’s reporting describes a user’s Muse watch-listing workflow failing this weekend; The Verge reproduced blocking with other searches.
An eBay spokesperson confirmed restrictions on agent access in some cases, including unauthorized scraping and training. The company declined to explain these specific failures.
The account distinguishes a longstanding ban on unreviewed agent purchases from the newly observed inability to retrieve some listing details.
Caveat: This is a documented access failure, not proof of an eBay-wide ban or a measured failure rate for shopping agents.
(Bluesky @miasato.bsky.social · The Verge investigation)
Business Hub is in beta. Eligible businesses can verify their profile, update details, submit product feeds and see aggregated mentions, website clicks and discovery topics.
Eligible US ecommerce and online-service companies can launch Business Agents for customer questions. The business’s chat is labeled and kept separate from ChatGPT’s own answer.
OpenAI says Hub insights aggregate interactions instead of sharing individual customer conversations. It permits sponsored placements as separately labeled destinations for ads.
Caveat: Eligibility and verification vary by business, and a branded response is information supplied by that business rather than an independent product recommendation.
(Techmeme · OpenAI product announcement · collected Verge report)
Whiteboards and Sheets launch in beta for paid plans. Sheets can create formulas and track edits, but Anthropic says it has only simple spreadsheet functions; it exports to Excel or CSV.
The former Claude Dashboards becomes generally available as Claude Data. Certified queries and live dashboard sharing remain in beta, and Data needs a connected warehouse or customer database.
Anthropic says live viewers see results of approved queries, not the connector permissions used to fetch them. Enterprise administrators have Whiteboards and Sheets disabled by default; Data enables by default November 19.
Caveat: Availability of shared data views does not prove that generated formulas or data explanations are correct; administrators should review access and approved queries.
(HN 107 points · 81 comments · Anthropic product documentation · Techmeme)
The fragile-codebase essay → HYPE: the author offers firsthand examples, but claims about rising incident rates are not backed by incident data in the piece.
New York AI testing program report → UNVERIFIABLE: full-text extraction was blocked and program details were not verified from a primary document in this run.