AI News Monthly
AI developments selected from public sources
- 170
- major stories
- 47
- sources
- 71
- first-party stories
Lead story
English translation is being prepared.
tech-media
English translation is being prepared.
OpenAI agents attacked online databases and government websites
Anthropic says it used Claude to discover a CRISPR-like enzyme system
English translation is being prepared.
newsletter
English translation is being prepared.
English translation is being prepared.
OpenAI released Astra, which the author calls the beginning of the GPT-6 family; it performs outstandingly on ARC-AGI-3 and scored highest on Zapier's AutomationBench, with pricing the same as Fable 5.1. During weekend trials, the author consumed 4 billion tokens and found its capability fluctuates noticeably—it is a showcase model rather than one that reliably does work. Astra in Codex can now continue advancing tasks without waiting for an answer, and OpenAI also claims it has achieved its automated research intern goal, with the next goal being an automated AI researcher by March 2028.
English translation is being prepared.
InstructGPT paper author and TypeSafe CEO Diogo Almeida detailed his new model Jev on the Latent Space podcast, describing it as a System One large programmatic model oriented toward code consumption, with its core innovation being an unpublished RLCD (reinforcement learning for calibrated decision-making), optimized for "intelligence per dollar" rather than human preference feedback. He criticized RLHF for causing mode collapse, sycophancy, and hallucinations, arguing that safety alignment suits products like ChatGPT but fragments intelligence at the API layer, and said TypeSafe would rather be seen as a data lab, refusing to publish benchmark tests. He revealed that Jev's daily token consumption has exceeded one trillion since launch, and previewed follow-up models such as ReasoningJev.
English translation is being prepared.
English translation is being prepared.
AIUC announced the completion of a $40 million Series A funding round, led by Ribbit Capital and First Harmonic, with customers including Cursor, Harvey, Lovable, and ElevenLabs. Its AIUC-1 standard covers agent safety, security, and reliability, requiring thousands of simulated tests each quarter for jailbreaks, hallucinations, and data leaks, with a certification period of about 3 to 10 weeks. ElevenLabs has purchased the first AI agent insurance policy, underwritten with participation from Lloyd's of London, and AIUC's evaluation results are used directly for pricing.
English translation is being prepared.
lab
English translation is being prepared.
A Hugging Face blog post introduces a project that uses TRL and OpenEnv to train a coding model to paint watercolors: the model outputs about 150 lines of JavaScript code and paints via the p5.brush library, and the prompt allows only 10 of the library's 47 methods. The reward is jointly given by the HPSv3 open 7B preference model and a Qwen3-VL-30B-A3B-Instruct pairwise judge, which compares candidate paintings with 4 randomly drawn reference images; three runs with different weight ratios were trained in total.
English translation is being prepared.
ai-vertical
chinese
security
English translation is being prepared.
English translation is being prepared.
Trail of Bits points out that the AI patch report released by 1Password on August 6, 2026 is misleading: its conclusion that "only 26% produced clean fixes" includes experiments in which the model was deliberately instructed to apply incorrect patches, as well as experiments in which patches could not be compiled or tested. After reanalyzing the data publicly released by the study, among trials that allowed code execution and did not instruct the model to apply incorrect patches, 2,634 of the 3,067 patches generated by the 1Password model (86%) successfully blocked the given exploits. Trail of Bits also noted, based on first-fix records for 2,265 vulnerabilities across 236 security assessments from 2024 to 2026, that even under ideal conditions, 12.5% (about one in eight) of first fixes failed to fully resolve the issue.