← 全部往期

Promptea Weekly

2026-11-01至2026-11-07这一周

本周头条

OpenAI releases GPT-6 Cyber, a gated exploit-hunting model it rates Critical for cyber risk

The GPT-6 Astra variant is API-only for approved Daybreak Red organizations, costs $20/$100 per million tokens, and comes with a system card full of vendor-run evaluations.

本周AI动态

security

OpenAI releases GPT-6 Cyber, a gated exploit-hunting model it rates Critical for cyber risk

The GPT-6 Astra variant is API-only for approved Daybreak Red organizations, costs $20/$100 per million tokens, and comes with a system card full of vendor-run evaluations.

It shows OpenAI's pattern for dual-use capability: a specialized, less-refusing model released only behind identity verification, organizational controls and monitoring rather than in general products.

阅读全文 →

model-release

Tencent launches Hy4 with vision and computer use, but the open weights are still to come

Tencent's flagship model gains image understanding and agentic computer control at the preview's price. The weights and system cards are promised "in the coming days".

A 770B-class MoE model with vision and computer use, priced under $1 per million input tokens, adds a low-cost option for multimodal agent work.

阅读全文 →

research

Ai2 releases ToyBlocks, and finds agents write better tool descriptions than experts

An open robot-learning simulator built on blocks and a single arm, plus a result any MCP developer can test: self-written skill descriptions beat expert ones.

Tool and skill descriptions are prompts the model reads on every call; Ai2's numbers suggest that having the agent probe a tool and document it can beat hand-written docs.

阅读全文 →

business

OpenAI and Meta buy coding-agent tooling on the same day

OpenAI agreed to acquire Pierre, the startup behind Code Storage, to strengthen Codex, and Meta bought Conductor, which runs coding agents in parallel. Neither disclosed a price.

The competition among AI labs is moving from models to the layers around them: repository storage, code review and multi-agent orchestration.

阅读全文 →

benchmarks

Ai2's StateBench finds frontier models lose track of changing values in long inputs

A new benchmark separates finding a fact from following it through a chain of updates, and frontier models fail the second as documents grow.

Needle-in-a-haystack style tests mostly measure retrieval, the easiest StateBench level, so a large advertised context window does not guarantee a model can follow values that change across a document.

阅读全文 →

developer-tools

Hugging Face proposes CPU-sized contracts to catch silent Transformers regressions

Model Integration Contracts pair shrunk model fixtures with human-reviewed baselines. In a replay, one flagged a UMT5 training bug that sat in main for almost two months.

Silent numerical regressions, where nothing crashes but a model trains or generates differently, are among the hardest bugs for anyone fine-tuning or serving open models to notice.

阅读全文 →

你可能错过的工具和发布

OpenAI releases GPT-6 Cyber, a gated exploit-hunting model… · Promptea