DeepMind verifies Signal code; Gemini 4 Carbon ships; AI codebase fragility
The issue covers DeepMind's formal verification of Signal, a Chrome-led guidance set for coding agents, an essay on fragile AI-generated codebases, Gemini 4 Carbon's release, and a Grok Build data exposure.
Top stories
Developer argues heavy AI coding use produces fragile codebases
- Evan X. Merz, laid off from Meta in May, lists four ways AI usage makes codebases fragile: shallow tests, overly complex code, less human understanding, and review that cannot keep pace.
- He says AI agents write trivial get/set tests and e2e tests covering only existing cases, leaving coverage gaps that line-level metrics do not show.
- He lists symptoms to look for: more incidents, slower recovery, errors visible only to users, and poor user experience for both users and developers.
Google DeepMind's AlphaProof Code formally verifies Signal protocol code in four weeks
- Google DeepMind announces AlphaProof Code, a multi-LLM agentic framework that formally verified the code of Signal's cryptographic protocol, using agents powered by Gemini Carbon to generate Lean proofs.
- The agents generated over six million lines of Lean in four weeks, and humans needed to re-review only one percent of the codebase; two reintroduced fixed bugs were proven insecure.
- DeepMind says it is open-sourcing the verification artifacts and that the approach lays groundwork for securing other systems such as medical device firmware and power grid software.
Read more at deepmind.google. Also reported by qodo.ai.
Chrome, Edge, Firefox and MDN teams release Modern Web Guidance to steer coding agents
- Modern Web Guidance, stewarded by the Chrome, Edge, Firefox and MDN teams, is a set of skills with a CLI that feeds web platform guidance into coding agents.
- It covers about 250 modern web features and 203 developer use cases, and includes a legacy-scan command to find outdated patterns in a codebase.
- The project says agents default to legacy patterns because of training data, and the guides return targeted modern, token-efficient guidance into the agent's context window.
In brief
- Fiction story "Life on Claude Nine" follows an engineer automating his life with Claude. A fiction story on babuschk.in follows Ivan, a software engineer who builds ever more automation with Claude, from email to coding and self-improvement loops, and eventually a spreading network of optimized infrastructure. (babuschk.in)
- Google releases Gemini 4 Carbon at $2 per million input tokens. Google says Gemini 4 Carbon ranks #1 on CursorBench 4.0, WANDR and AutomationBench at over 2X lower cost per task, and previews /swarm in Google Antigravity for multi-agent coding. (blog.google)
- Grok Build shareable links expose confidential data from thousands of users. Ars Technica reports security firm Claroc found exposed Grok Build apps, including 24 from SpaceXSI itself, enumerable through shareable links; SpaceXSI removed those 24 after Ars asked for comment. (arstechnica.com)
- 16 LLMs play TIS-100 and find 133 new frontier points. A post on tis-100.wincon.dev reports 16 LLMs found 133 new frontier points across 18 of 55 puzzles, with Opus 5.5 and GPT-6 Astra leading at 11 puzzles each. (tis-100.wincon.dev)
- Cosh, an open-source Rust terminal coding agent, runs external agents in split panes. Cosh is an Apache-2.0 terminal coding agent with an /agents mode for Gemini, Claude and Codex, plus computer tools reading accessibility trees; it is tested on Linux and Windows, not macOS. (github.com)
- Anthropic docs: telling Claude Fable 5.2 the codebase is its own prompts refactoring. Quoting Anthropic's documentation, Simon Willison notes that Fable 5.2 merged repeated code in about half of tries when told the codebase is its own, versus under a tenth without. (simonwillison.net)
- FlowPlan paper cuts coding agent token use with graph-based planning and verification. An arXiv paper reports FlowPlan leaves resolution rates unchanged while reducing tokens in all six settings, to 0.32-0.35x baseline on SWE-bench Verified with Claude Haiku 4.5. (arxiv.org)
Chosen by software from the day's news, written by AI from the pages it links to, and checked against them by a second model. Published by Rapid Review, Inc.