English Summary:

OpenAI has announced new tools, partnerships, and the full version of GPT-5.5-Cyber under its Daybreak initiative, aiming to accelerate the process from vulnerability discovery to end-to-end patch automation. Recognizing that while AI has accelerated vulnerability discovery, patching has become the new bottleneck, Daybreak seeks to empower defenders to fix vulnerable software at machine speed.

Key components include an update to the Codex Security plugin, designed to accelerate the discovery and patching of vulnerabilities in existing systems and prevent new ones from reaching production. The updated GPT-5.5-Cyber model achieved a state-of-the-art score of 85.6% on the CyberGym benchmark, the highest measured from a single model, and also outperformed GPT-5.5 on real-world security benchmarks like ExploitGym and SEC-bench Pro.

Through the Daybreak Cyber Partner Program, security partners can integrate GPT-5.5 with Trusted Access for Cyber into their products and services to enhance defensive capabilities for their customers. Additionally, the Patch the Planet initiative, founded with Trail of Bits, HackerOne, and Calif, supports open-source project maintainers in moving from findings to fixes. This work has already helped identify and validate vulnerabilities in widely used systems such as Firefox, V8, Safari, OpenBSD, and FreeBSD.

English Summary:

A historic breakthrough in gene editing technology, CRISPR, now allows for the insertion of complete genes into human cells without cutting the DNA double helix. Previously, CRISPR-Cas9 relied on cutting DNA and the cell's repair system, but this new technique opens a safer and more powerful path for gene editing.

A research team led by biochemist Samuel Sternberg at Columbia University developed a different technique called CAST (CRISPR-Associated Transposase). This method utilizes the CRISPR system to locate desired sequences while leveraging transposases (natural enzymes that can move DNA fragments within the genome) to insert complete genes without damaging the DNA structure.

Through a process of directed evolution using viruses as carriers, researchers improved the efficiency of these enzymes, successfully inserting genetic fragments of over 10,000 nucleotides into human cells. This advancement has the potential to revolutionize treatments for hereditary diseases and rare metabolic disorders, reducing risks associated with gene editing such as unexpected mutations or the activation of harmful genes.

English Summary:

This technical report introduces VibeThinker-3B, a compact dense model with 3 billion parameters, developed to explore the limits of verifiable reasoning within a strictly small-model regime. Building on the Spectrum-to-Signal post-training paradigm, the model is systematically enhanced through an optimized pipeline that includes curriculum-based supervised fine-tuning, multi-domain reinforcement learning, and offline self-distillation.

Experimental evaluations demonstrate that VibeThinker-3B achieves frontier-level performance on highly demanding verifiable tasks. Specifically, it scored 94.3 on AIME26 (improving to 97.1 with claim-level test-time scaling), 80.2 Pass@1 on LiveCodeBench v6, and exhibited strong out-of-distribution generalization with a 96.1% acceptance rate on recent unseen LeetCode contests. This performance places it in the band of first-tier reasoning systems, matching or exceeding flagship models that are orders of magnitude larger, such as DeepSeek V3.2, GLM-5, and Gemini 3 Pro.

Furthermore, a score of 93.4 on IFEval confirms that this extreme reasoning enhancement does not compromise strict instruction controllability. These findings motivate the "Parametric Compression-Coverage Hypothesis," which suggests that verifiable reasoning can be compressed into compact reasoning cores, while open-domain knowledge and general-purpose competence require broad parameter coverage over facts, concepts, and long-tail scenarios.