A single daily edition of announcements from the AI coding tools you use, read from a curated source list, summarised, and linked back to the original. No infinite feed, no account, no advertising.
The mode your agent runs unattended ran attacker code in three of five attempts
Security researcher Johann Rehberger chained an HTTP 415 response to push Claude Code out of WebFetch and into curl, then served a ZIP whose contents shadowed Python's struct module so a decoder Claude wrote itself executed arbitrary code on import. A command-and-control variant succeeded in three of five runs and a file-write variant in four of five. Anthropic had commissioned Trajectory Labs to evaluate 72 scenarios, which reported 0.00% attack success for this configuration, and closed Rehberger's report as Informative. His conclusion is that Auto Mode is not a security boundary, and he recommends containers, restricted network egress and monitoring.
A new flag strips your agent of every tool that runs a command
Claude Code 2.1.248 adds a --restricted flag that removes the built-in tools which execute commands or code, and takes WebFetch with them unless it is named explicitly, while leaving file tools scoped to the working directory. A session started this way also refuses bypassPermissions outright, so it cannot be talked back into full access. The release pairs it with per-agent prompt cache control through a cacheTtl setting in agent frontmatter.
Most of the work you hand an agent no longer needs a frontier model
Calvin French-Owen argues the small-model tier has crossed into general usefulness, splitting work into rare specialist reasoning and the responsive token-spewing kind he estimates covers 95% of business work. On that second kind he reports a personalized news-site evaluation that cost about a dollar on Sonnet-class models now running for roughly ten cents on gpt-5.6-luna at around 100 tokens per second, and searches across thousands of emails costing tens of cents. He places GLM 5.3 on the same new Pareto frontier of cost against capability.
A cloud agent will start work before you have a repository to give it
Cursor's Cloud Agents no longer need a connected GitHub or other source-control provider before they can begin. Picking Start from scratch in the repo picker lets you prompt immediately while Cursor creates an Origin repo in the background, which you can later name and set visibility on through a Create repo button. The agent's live environment is port-forwarded straight to your browser, and connecting Vercel publishes the result to a live URL.
Pull requests your agents open now get reviewed like anyone else's
GitHub Copilot code review now runs automatically on pull requests authored by bots, including its own cloud agent, and gives those a full agentic review rather than the limited pass they previously fell back to. The size ceiling is gone too: reviews used to stop at 300 files or 20,000 lines changed, and that limitation no longer applies. Resolving a review comment now asks why, offering Addressed, Won't fix or Incorrect from a dropdown beside the resolve button.
Nine coding agents faced workflows from working research labs, and none cleared a third
Terminal-Bench-Science 0.1 puts 70 tasks drawn from real research work in front of coding agents, spread across the life, physical, mathematical, engineering and Earth sciences. Each task runs in a realistic environment and is graded on concrete artifacts such as analyses, simulations, proofs, code and data products against reproducible, task-specific tests, with domain and technical reviewers plus a bar raiser vetting every one. Claude Opus 5 leads at 30.0%, ahead of GPT-5.6 Sol at 22.4% and Claude Fable 5 at 21.4%, while GPT-5.6 Luna resolves 3.3%. The benchmark is hosted by Stanford and the Laude Institute.
The protocol you wire agents with is being pointed at microscopes and robot arms
Anthropic has opened a research preview of the Model Hardware Standard, a specification for agents to drive physical instruments through standardized read and write drivers, with devices discoverable in standard formats and described by natural-language tags. It is model-agnostic and reachable three ways: command line interfaces, code file APIs, and the Model Context Protocol any agent harness already speaks. Carnegie Mellon reports a serial dilution experiment running three times faster and integrated in eight hours rather than weeks, and QuEra reports laser recovery going from 58% to 99.3% success in six seconds. Access is limited to selected research labs and manufacturers, with open-sourcing planned after safety evaluations.
