{"slug": "vercel-s-1m-sandbox-challenge-finds-a-kernel-zero-day-reaching-the-host", "title": "Vercel's $1M sandbox challenge finds a kernel zero-day reaching the host", "summary": "Vercel's $1 million sandbox escape challenge drew 1,285 vulnerability reports and paid out $242,000 in bounties, with the top finding a Linux kernel networking flaw that let a microVM guest leak host memory, crash the host, or run code, though no customer data was accessed. The challenge results were reported alongside other AI engineering items, including Anthropic exporting OpenTelemetry traces from the Messages API, Managed Agents sessions, and the Claude SDKs, and Siemens EDA raising a chip-design agent's benchmark score from about 55% to 95% using NVIDIA NeMo Gym verifiers.", "body_md": "Vercel's sandbox escape challenge drew 1,285 reports and paid $242,000 in bounties. The top find was a Linux kernel networking flaw that let a microVM guest leak host memory, crash the host or run code, with no customer data accessed.\nRead: Vercel's sandbox escape challenge drew 1,285 reports and paid $242,000 in bounties. The top find was a Linux kernel networking flaw that let a microVM guest leak host memory, crash the host or run code, with no customer data accessed.\nRead: Anthropic now exports OpenTelemetry traces from the Messages API, Managed Agents sessions and the Claude SDKs, and LangSmith accepts them on day one.\nRead: SemiAnalysis says Google submitted TPUv7 FP8 vLLM results to its Agentic Inference Benchmark, showing 1.67x better performance per dollar than GB300 NVL72 at FP4.\nRead: Subframe's Real Design Benchmark asks people to choose between a real product screen and one built by Opus 5.5 or GPT-6.1 Sol. After thousands of votes, 57% preferred the AI design.\nRead: Epoch AI's SpireClimb has models play Slay the Spire through the game's graphical interface. Frontier models now win over 90% of easy runs and 20% at the hardest difficulty.\nRead: Siemens EDA used NVIDIA NeMo Gym verifiers to find failure modes in a chip-design agent, then rewrote its skills and MCP tool descriptions. Its benchmark score rose from about 55% to 95% as turns fell from 155 to 53.\nRead: SWEeper-Bench gives coding agents a codebase and an open-ended request, so they must write interactive tests to find bugs in 200 real web apps. The best of 15 agents passes 59%.", "url": "https://wpnews.pro/news/vercel-s-1m-sandbox-challenge-finds-a-kernel-zero-day-reaching-the-host", "canonical_source": "https://www.vibeleaderboard.ai/intel/brief/2026-10-15", "published_at": "2026-10-15 11:16:48+00:00", "updated_at": "2026-10-15 11:47:41.777695+00:00", "lang": "en", "topics": ["ai-safety", "ai-agents", "ai-infrastructure", "developer-tools"], "entities": ["Vercel", "Linux", "Anthropic", "Claude", "LangSmith", "Siemens EDA", "NVIDIA", "NeMo Gym"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/vercel-s-1m-sandbox-challenge-finds-a-kernel-zero-day-reaching-the-host", "markdown": "https://wpnews.pro/news/vercel-s-1m-sandbox-challenge-finds-a-kernel-zero-day-reaching-the-host.md", "text": "https://wpnews.pro/news/vercel-s-1m-sandbox-challenge-finds-a-kernel-zero-day-reaching-the-host.txt", "jsonld": "https://wpnews.pro/news/vercel-s-1m-sandbox-challenge-finds-a-kernel-zero-day-reaching-the-host.jsonld"}}