October 2026 Summaries

Vercel’s two-week $1 million HackerOne challenge for Vercel Sandbox received 1,285 reports, validated 57 findings, and paid $242,000 in bounties, while an independent Trail of Bits review identified 20 additional findings and judged the core Firecracker microVM virtualization boundary sound. The most serious report uncovered an upstream Linux kernel networking zero-day that could leak host memory, crash hosts, or potentially enable host code execution under demonstrated conditions; Vercel coordinated with the researcher, AWS, kernel maintainers, and other partners on mitigations and pending patches. No report demonstrated access to customer files, secrets, or application data. Other findings included a device-mapper issue that could expose recycled disk blocks and a snapshot restore trust-boundary flaw that allowed two accounts controlled by the same researcher to swap and access each other’s snapshots, although targeting an unrelated tenant required an undisclosed random storage identifier. Vercel said it addressed these issues by reducing trust in guest-provided data, centralizing network-policy enforcement, adding defenses for potential future virtualization flaws, and moving snapshot management into host-side, project-scoped services.

Jev is a new AI model designed as a universal classifier that rapidly makes structured, narrow decisions by answering multiple-choice, scoring, or probability questions with confidence estimates rather than generating free-form text. Vercel’s AI SDK for Python adds experimental support for Jev through an evaluate() API, allowing developers to provide model state and typed questions using an AI Gateway key. The post describes Jev as an LLM adapted to classifier-like behavior, offering speed, low cost, JSON outputs, and broad pretrained knowledge without task-specific model training, while emphasizing that developers must independently evaluate its accuracy. In one practical experiment, Jev classified partially typed inputs as Python or English more effectively than a custom-trained classifier but still made ambiguous mistakes. A second, intentionally impractical experiment attempted to make Jev generate Python code through successive classification choices to construct an abstract syntax tree; it produced syntactically valid but often incorrect code, illustrating the difficulty of using a classifier as a text generator. The author encourages developers to experiment with Jev using Python 3.12+, the AI SDK, and a simple interactive classifier playground.