OpenAI's Hacker Model Answers 95% of Requests Without Refusal; Claude Advances Riemann Hypothesis Record

OpenAI Releases a "Hacker Model": A Double-Edged Sword for Defenders

In the field of AI security, a long-standing paradox exists: the more powerful a model becomes, the more it can be used for both attacks and defense. OpenAI's recently launched GPT-5.6 Cyber, part of a project called Daybreak Red, has formally brought this contradiction into the open.

Reportedly, this model is only available to vetted defenders. In OpenAI's internal evaluations, it completes 95% of advanced cybersecurity requests. A critical distinction needs clarification here: 95% refers to the response rate, not the attack success rate. In other words, the model no longer refuses to answer high-risk security questions across the board as it did in the past — instead, it chooses to assist with analysis and provide solutions, provided the user's identity is verified.

From Refusal to Assistance: Redefining Security Boundaries

The model also helped fix a critical vulnerability in Chrome's V8 engine. V8 is Google's high-performance JavaScript and WebAssembly engine, written in C++. It not only powers the Chrome browser but also serves as the core runtime for Node.js. Given JavaScript's ubiquity in the modern web, any vulnerability in V8 — especially memory safety issues like type confusion or out-of-bounds read/write — could be exploited by attackers to achieve remote code execution (RCE), directly threatening over 3 billion Chrome users worldwide. Historically, V8 vulnerabilities have been repeatedly used in Advanced Persistent Threat (APT) attacks and zero-day exploit chains, making them among the highest-bounty targets in national vulnerability reward programs. This case demonstrates that when powerful offensive and defensive capabilities are in the hands of defenders, they can accelerate the process of vulnerability discovery and patching.

But controversy follows: should such capabilities be opened up or restricted? If the vetting mechanism is bypassed, or if there are loopholes in determining "defender" status, this system could become an attacker's weapon. OpenAI's "defenders-only" approach essentially places the security burden on access vetting — this is both a technical and a governance challenge.

Claude Advances a Record Related to the Riemann Hypothesis: The Real Boundaries of AI in Mathematics

The second piece of news is equally significant. Anthropic's Claude pushed an existing record related to the Riemann Hypothesis from 41.6% to 67.2%. It must be specifically emphasized: Claude did not prove the Riemann Hypothesis — it only advanced a related mathematical record.

The Riemann Hypothesis was proposed by German mathematician Bernhard Riemann in 1859, asserting that all non-trivial zeros of the Riemann ζ function have a real part equal to 1/2. It is one of the seven Millennium Prize Problems designated by the Clay Mathematics Institute, with a $1 million bounty still unclaimed. The "related record" typically refers to numerical or analytical advances closely connected to the distribution of ζ function zeros, or upper/lower bound estimates of the De Bruijn-Newman constant Λ. In 2018, Rodgers and Tao proved that Λ≥0, and prior records have been continuously pushed forward. Claude's improvement from 41.6% to 67.2% belongs to this category of numerical advances in boundary estimation — a step toward the conjecture's direction, but still fundamentally far from a complete proof.

31 Million Tokens and 60 Sub-Agents: The Computational Investment

The computational investment behind this achievement is quite staggering. The entire process was conducted in two rounds, with Claude Code outputting a total of 31 million tokens. The second round deployed approximately 60 sub-agents, running continuously for a day and a half.

Sub-agents represent an architectural pattern where complex tasks are decomposed and processed in parallel by multiple specialized AI instances. In this mathematical endeavor, approximately 60 sub-agents each handled different computational path explorations, lemma verifications, or numerical experiments, with a master agent aggregating results and making global judgments. This pattern draws from the collaborative approach of human mathematician teams: one person handles algebraic transformations, another verifies numerical results, and a third checks boundary conditions. The 31 million token output is equivalent to approximately 50-70 academic monographs in text volume, indicating that AI requires extensive "thinking process" unfolding during long-range reasoning to approach reliable conclusions. This "multi-agent collaboration + extended reasoning" paradigm is becoming the new model for AI to tackle complex problems.

Scientific Rigor: Verification Is Not Yet Complete

Currently, this result has been verified by internal company mathematicians, and external experts have conducted brief reviews. But — and this is the most critical qualification — it has not yet undergone standard peer review.

Peer review in mathematics differs from other disciplines — it demands that every step in the logical chain withstand completely formal verification, and a single minor error can cause the entire conclusion to collapse. History is replete with cases where initial claims of major breakthroughs were found to contain errors during review, such as Shinichi Mochizuki's 2012 proof of the ABC conjecture, which remains controversial to this day. For AI-assisted mathematical work, review additionally requires verifying whether "hallucinations" exist in the AI's reasoning process — situations where the surface logic appears coherent but actual steps are erroneous. Therefore, before peer review is completed, any numerical record can only be regarded as a "candidate result pending verification."

In an era where AI routinely "announces major breakthroughs," this kind of restrained expression is actually precious. The mathematics community demands extremely high standards of rigor, and any result that hasn't undergone peer review cannot be considered definitive. AI's ability to advance records is one thing; whether those advances can withstand scrutiny from the mathematical community is another.

Meta Open-Sources Muse Glimmer: A 30-Billion Parameter Local Agent Model

The third piece of news comes from Meta. The company has released the weights for its Muse Glimmer model with 30 billion parameters under an Apache 2.0 license, targeting code generation, tool calling, and local agent scenarios.

Apache 2.0 is currently one of the most permissive mainstream open-source licenses, allowing commercial use, modification, and redistribution without requiring derivative works to be open-sourced (unlike GPL). However, an important distinction exists between "open weights" and "fully open source": the former only releases the trained model parameter files, while the latter should also include training code, datasets, data processing pipelines, and training hyperparameters. Meta's choice to open weights without releasing training data allows the community to fine-tune and deploy the model while protecting its data assets and training methodology. This "semi-open" strategy is becoming the industry's mainstream approach.

Runs on a Single Mac or Consumer-Grade GPU

The most striking aspect is its deployment threshold: it can run on a single Mac or a single consumer-grade GPU. A 30-billion parameter model requires approximately 60GB of VRAM at FP16 precision, but through quantization techniques (such as 4-bit GPTQ or GGUF format), it can be compressed to approximately 15-20GB — just fitting into Apple M-series chips' unified memory (starting at 32GB with M2 Max) or a single RTX 4090's 24GB VRAM. Key technologies include: Group Quantization that compresses models while maintaining inference quality; Flash Attention that reduces memory bandwidth bottlenecks; and Apple's MLX framework with deep optimization for Metal GPU. This allows individual developers to run large models with tool-calling and agent capabilities on desktop-class devices for the first time.

However, two clarifications are needed: first, this isn't the kind of "small model" designed for phones; second, open weights don't mean all training data is open-sourced. The real takeaway is — personal AI assistants can now go offline and stay on your own device.

For users concerned about privacy and data sovereignty, this is an important signal. When powerful agents can run entirely locally, data no longer needs to be uploaded to the cloud, opening new possibilities for AI applications in sensitive scenarios.

Tencent WorldCloud: Generate Explorable 3D Worlds from a Single Sentence

The fourth piece of news comes from China. Tencent's team has released the WorldCloud project: a single prompt is decomposed into terrain, regions, and assets, then a Blender agent performs self-checks, ultimately generating an explorable, editable, layered 3D world.

Stunning Visuals, But Still at the Research Demo Stage

From the demo footage, the results are quite stunning — automated generation from text to complete 3D worlds represents the future direction of content creation tools. But we need to remain level-headed: currently it's more of a research demonstration, and the repository hasn't yet provided complete code and weights.

Additionally, the experimental process also called upon external models like Claude and GPT Image. WorldCloud represents an "AI Orchestration" paradigm: it isn't a single generative model itself, but rather a task scheduling and quality control framework. The workflow typically proceeds as follows: a large language model (such as Claude) understands natural language instructions and plans the world structure; an image generation model (such as GPT Image) creates textures and concept art; a 3D generation model converts 2D concepts into three-dimensional assets; and finally a Blender agent serves as the execution layer, completing assembly, lighting, and physics property setup while self-checking to ensure scene consistency. The advantage of this pipeline design is that each component can be upgraded independently; the disadvantage is high end-to-end latency and dependence on external API stability. This is also a common characteristic of many "stunning demos" today: the innovation lies more in novel technical combinations than in fundamental breakthroughs.

The AI Deployment Trends Behind These Four Stories

Looking at these four stories together, several clear threads in AI development emerge:

• OpenAI is exploring the boundaries of opening up security capabilities, balancing offense and defense through access vetting;
• Anthropic is demonstrating the potential of multi-agent collaboration to tackle scientific challenges while maintaining scientific rigor;
• Meta is pushing powerful models toward localization and decentralization;
• Tencent is exploring the possibilities of multi-model orchestration for generating complex 3D content.

They respectively correspond to four directions: security governance, scientific research, privacy-preserving deployment, and content generation. And one consensus runs throughout: In an era where AI capabilities are advancing at breakneck speed, accurately describing results and clearly understanding boundaries is more important than "breakthroughs" alone.

Response rate is not attack success rate; advancing a record is not proving a conjecture; open weights is not open-source data; a stunning demo is not a usable product — these distinctions are precisely the key to understanding AI's real progress.