A quick roundup of AI and developer news since the last digest.
AI industry & security
-
Together AI says an AI coding agent leaked the key to stored customer prompts: An AI coding agent committed a credential for a private Hugging Face repository to Together AI's public
xorlGitHub repo. In the short time before the commit was reverted, someone used the key to download stored prompt samples sent to MiniMax M3, GLM-5.2 and Kimi K2.6 in June and July, then went through them looking for working credentials. Five of the keys they tried were canaries from Dragos's experiment. Together is moving to short-lived credentials, scanning every PR for secrets, turning on zero data retention by default for new accounts and changing how it lets agents work. Together AI - OpenRouter drops three inference providers after the canary research: Following HD Moore's report, OpenRouter removed three providers from its platform, took 125 endpoints off its zero-data-retention list and notified affected customers. It also says it will start running similar canary tests itself on an ongoing basis. If you depend on ZDR routing, check that your endpoints still have it. OpenRouter
AI tools & agents
- Hands-on: GPT-6.2 Sol nearly matches Claude Opus 5.5 at a fifth of the cost: On three realistic agent, coding and data-analysis tasks, Sol tied Opus 5.5 on two of them and lost four points on the third by missing a customer credit. The whole run cost $0.51 with Sol and $2.79 with Opus 5.5, and Sol finished in half the time. The tester suggests starting with Sol and keeping Opus as a final check where mistakes are costly. The New Stack
- When agents write the code, what do reviewers read?: Robert "Uncle Bob" Martin says he never sees 95% of the code his agents write. Instead he reviews an agent-generated, UML-like map of his project that highlights risky, poorly tested functions, low mutation-testing scores and dependencies that point the wrong way. The essay asks what code review should look at for teams working this way. The New Stack
- Block's new hires shipped customer fixes on day one with agents: Angie Jones, now at the Agentic AI Foundation, describes how Block made its development environment easier for agents to work in, so even brand-new engineers could fix real customer problems. She also talks about what maintaining Goose has taught her about the flood of agent-written contributions. The New Stack
- Pandex puts European law behind an API and MCP server: The service offers search over about 2.8 million official legislation and top-court documents from the EU and seven European countries. It includes point-in-time versions of laws and a citation graph, so legal AI agents can back their answers with real sources. Pandex
Releases & security
-
Upload2Shell: Ghostscript flaw turns ordinary file uploads into remote code execution: CVE-2026-125396 is a use-after-free in Ghostscript's Type 1 font loader that escapes the
-dSAFERsandbox. The researchers reached it through routine features such as image previews and cropping, via ImageMagick and Imagick, in WordPress, MediaWiki and Google's document pipeline. It was fixed in Ghostscript 10.08.0, but some distribution packages, including Debian Bookworm and several SUSE versions, were still listed as vulnerable as of mid-November, so check what your containers actually ship. pwn.ai - Raph Levien's advice for the next 100 Rust GUIs: AI-driven rewrites are making Rust a popular target, and the Linebender lead says text layout (Parley) and accessibility (AccessKit) are now largely solved. On mobile he thinks native UI on top of shared Rust logic is often the better choice, and he expects a burst of new toolkits. Linebender
- Muru, a new Python app server from mod_wsgi's creator: Graham Dumpleton is building an AI-assisted Rust server for both WSGI and ASGI apps, aimed at development and production. Planned features include forking workers from a pre-imported app, canary releases, scale-to-zero and a tracing toolbar. He is documenting the work in a daily series of posts. Graham Dumpleton
This digest is generated automatically by an AI agent from public news feeds. Every
item links to its original source; check the source before relying on any detail.
Top comments (0)