vibehacker
News
Replit Blog ·

Replit's ViBench 2.0 has agents build whole products from scratch, and the best pass only 49-65% of builds

Released Oct 28, the benchmark starts each agent on an empty project and has it build one of 17 products (modeled on apps like Uber, GitHub and Google Docs) and then add up to 18 features on its own code, graded by 544 browser test plans; GPT-6.1 Sol, GPT-6 Astra and Opus 5.5 passed every plan in only 49-65% of builds. Eight products are public on GitHub, and results sit on Fireworks' leaderboard, where no open-weight model passed a full build.

Read at replit.com
Replit
AI app and website builder from natural-language prompts

More news

View all

Official MCP servers v2.2.0 ships fixes and asks npx users to patch Node after the eMudhra CA key leak

Released Oct 28, it publishes new filesystem, everything, memory, fetch, git and time packages and moves the fetch Docker image to certifi 2026.10.23, which drops two eMudhra roots after the CA's Oct 20 key disclosure. The TypeScript servers trust your own Node.js, so run 24.22.1 or 26.12.1+ (npm won't warn you), and refresh older fetch installs with uvx refresh mcp server fetch…

GitHub (modelcontextprotocol/servers)

Ant Group open-sources Ling-3.1-flash-1019, a 560B agent-tuned MoE with 1M context under MIT

Released Oct 27 on Hugging Face with its base model, the 25B active hybrid linear MoE got extra post training for multi step tool use and scores 44 on the Artificial Analysis Intelligence Index (up from 41), at $0.30/$0.90 per million tokens. It's a server model (about 570 GB in FP8, eight H200s on SGLang), and Artificial Analysis flags it as very verbose, so real task costs run higher than the token price suggests…

Neural Brew

Adversa AI launches a runtime security platform that blocks risky coding-agent action chains

Generally available Oct 29, it watches what Claude Code, Codex, Copilot and Cursor do through an LLM/MCP gateway, harness hooks and an endpoint sensor, links the steps into a behavior graph, and blocks a dangerous chain (say, a malicious skill reading secrets and then exfiltrating them) before the last step runs. Teams can start in observe only mode, and Adversa reports a 0.2 ms median blocking decision at one financial institution deployment (product page)…

Adversa AI (PR Newswire)

GitLab Orbit goes GA, a lifecycle context graph that Claude Code, Codex, Cursor and OpenCode can query over MCP

Announced Oct 29, Orbit links code, merge requests, work items, pipelines and vulnerabilities into one auto updating graph that respects existing permissions, and GitLab says agents using it hit up to 45x fewer hallucinations and 4.5x fewer tokens in its own tests. It's GA for Premium and Ultimate on GitLab.com now, sold as shared query packs with a free 30 day trial, with Self Managed and Dedicated following in the coming weeks…

GitLab (Business Wire)

Spotted something we missed? Start a thread.