GPT-6 Cyber Breaks New Ground in Vulnerability Discovery, Gemini Video Editable at 15 Seconds, Ernie Open-Sources Model

The AI world saw multiple major updates overnight: OpenAI released GPT-6 Cyber, a specialized cybersecurity model with impressive vulnerability discovery capabilities; Google launched Gemini Omni 1.2 Flash with longer video duration and more flexible editing; Baidu open-sourced the Early 5.5 FlyerFlashbase foundation model; Anthropic rolled out personal finance features; ChatGPT's image generation was upgraded to Images 2.6; and DeepSeek Harness surpassed 1.54 million daily active users.

OpenAI Releases GPT-6 Cyber: Focused on Vulnerability Discovery and Remediation

According to the Bilibili AI news channel "Grok大小姐秘密花园," OpenAI has launched GPT-6 Cyber, a cybersecurity-specific model intended for authorized security research. The model can automatically find vulnerabilities, verify issues, and assist in writing patches. OpenAI emphasizes that the model tends to operate within the boundaries set by researchers, avoiding overreach.

The actual test data comes from official sources: in the Firefox browser, GPT-6 Cyber found 81 previously unknown high-severity vulnerabilities, which Mozilla is already fixing; on the Linux kernel, it reproduced 1,724 issues and produced 948 usable privilege-escalation validations. In the Exploit Gym benchmark, its success rate rose from 32.1% to 48.5%. Under OpenAI's own preparedness framework, its cybersecurity capability is classified as Critical. Currently, the model is only available via API and Codex to a select group of approved Daybreak Red customers, and is not yet accessible to general users.

This progress means AI's application in security research has entered a new phase, but it also raises concerns about dual-use risks. OpenAI balances innovation and safety through strict access controls.

Dual-use Risk: Refers to the same technology being usable for legitimate, beneficial purposes as well as being misused for harmful purposes. In cybersecurity, vulnerability discovery technology is both a core means for defenders to find and fix system weaknesses and potentially an entry point for attackers to find exploitable gaps. OpenAI classifies GPT-6 Cyber's cybersecurity capability as "Critical" in its preparedness framework, meaning that without restrictions, the model's ability to autonomously discover and exploit vulnerabilities could pose serious real-world threats. Therefore, OpenAI has adopted strict access controls, opening the model only to vetted "Daybreak Red" customers. This is a typical governance strategy for addressing dual-use risk—while promoting innovation in security research, it reduces the likelihood of misuse through identity verification, use commitments, and technical guardrails. Such risk governance frameworks are gradually becoming a standard evaluation step before the release of frontier AI models.

Google Gemini Omni 1.2 Flash: Longer, More Flexible Video Generation

Google released Gemini Omni 1.2 Flash, extending native video generation from 10 seconds to up to 15 seconds and allowing users to select clips from 3 to 15 seconds on a per-second basis. Users can hold multiple rounds of dialogue on the same video to modify the background, lighting, or aspect ratio without regenerating the entire clip.

The model supports 720P resolution and is priced at $0.07 per second, 30% lower than Omni 1.1. Users can first draft in 360P or 720P, then upscale the final output to 1080P or 4K. It is now generally available on the Gemini API and Google Cloud, and is rolling out on Flow, Google Width, and the Gemini App for Ultra users. Generated videos carry a SynthID watermark. Google acknowledges that fast sports motion and complex physics still need further optimization.

15-second clips with price-reduced revisions represent a significant cost and efficiency optimization for content creators. However, physical realism remains a weakness, making it suitable for marketing and social scenarios that demand real-time responsiveness.

ChatGPT Images 2.6: Faster Image Generation with GIF Support

ChatGPT's image generation capability has been upgraded to Images 2.6, now available on ChatGPT, Codex, and the API. OpenAI says generation is 50% faster than before. In the API, it is available in two versions, Flare and Sunburst, with Flare producing images in about 6 seconds at low quality. New capabilities include directly creating GIFs, with adjustable speed and frame selection. It is now available on desktop, mobile, and web across all subscription tiers.

OpenAI also disclosed usage data: about 4 billion images are generated per week, averaging over 500 million per day. Faster generation and GIF support further lower the barrier to creative expression, especially for social media content creation.

Anthropic Claude Launches Personal Finance Features

Starting this week, Anthropic is gradually rolling out Personal Finance in beta to Pro, Max, and Ultra users in the United States. Users can connect their chosen bank and credit card accounts via Plaid and ask about their financial situation in natural language.

The feature is read-only across the board—it cannot transfer funds, manage accounts, or invest. Finance conversations are not used to train models, and there are no ads or sponsored recommendations. The free tier, Team, and Enterprise versions are not yet supported, and it is limited to U.S. accounts. This restrained design emphasizes privacy and security but limits the range of use cases.

Plaid: Plaid is a fintech company that provides application programming interfaces (APIs) enabling third-party apps to securely connect to users' bank accounts, credit cards, and other financial accounts. After users complete identity verification through Plaid, apps can access transaction records, balances, account types, and other information in a read-only manner without directly touching users' bank login credentials. Plaid covers the vast majority of mainstream financial institutions in North America and is a commonly used underlying connection service for personal finance, payment, and lending apps. In Anthropic's personal finance feature, Plaid serves as a secure bridge between users' banks and Claude, ensuring that data access is limited in scope and auditable. This complements the feature's restrained design of being "read-only across the board and unable to transfer funds," safeguarding user funds and privacy at the technical level.

Baidu Open-Sources Early 5.5 FlyerFlashbase Foundation Model

On November 5, Baidu open-sourced Early 5.5 FlyerFlashbase, with a total of 28 billion parameters and only 1.4 billion activated per token. The model was pretrained from scratch on 33.5T tokens, with a maximum training context of 128K and support for up to 1M during fine-tuning and inference.

Note that these are the weights at the end of pretraining, without mid-training or post-training, making them suitable for researchers to fine-tune themselves. Baidu says the post-trained version of Early 5.5 Flash will follow. FP8 and BF16 foundation weights are available for download on Hugging Face. Releasing the foundation first and queuing post-training later is a clear strategy that provides the community with a research starting point.

Base Model and Post-training: A base model refers to model weights obtained through self-supervised pretraining from scratch on large-scale unlabeled or weakly labeled corpora (e.g., predicting the next token). It possesses general language modeling capabilities but has not yet been optimized for specific tasks (such as dialogue, instruction following, or safety alignment). Post-training builds on the base model using techniques such as supervised fine-tuning (SFT), reinforcement learning from human feedback (RLHF), or direct preference optimization (DPO) to teach the model to follow instructions, align with human preferences, and improve performance on specific tasks. Baidu's open-sourced Early 5.5 FlyerFlashbase is a "pretraining endpoint" foundation weight without mid-training or post-training, making it more suitable as a starting point for researchers to customize via fine-tuning rather than for direct deployment as a conversational assistant. This open-source strategy of "releasing the foundation first, then the post-trained version" both meets the research community's need for underlying weights and leaves room for the official release of a fine-tuned version later.

DeepSeek Harness Hits 1.54 Million Daily Active Users, Ecosystem Takes Shape

DeepSeek^Ad Harness officially announced that as of November 4, based on users of the official API, daily active users reached 1.54 million, with about 92 trillion tokens consumed per day. There are approximately 35,000 third-party plugins, and about 60% of active users have installed at least one plugin.

DeepSeek says it will next build an official plugin marketplace and prepare a support program for plugin authors. Harness launched V0.1 in August and released a desktop version in late September, with daily active users growing rapidly. The establishment of a plugin marketplace is expected to further prosper the ecosystem.

These six news items cover security, video, images, finance, open source, and ecosystem, showing that AI technology is accelerating its penetration into various industries.

One-line Summary

Six major AI updates arrived overnight, covering security, video, images, finance, open source, and ecosystem.

Paragraph Summary

This article summarizes six important AI updates from overnight. OpenAI released GPT-6 Cyber, a specialized cybersecurity model that performed notably in vulnerability discovery in Firefox and the Linux kernel, but is limited to authorized research use; Google launched Gemini Omni 1.2 Flash, extending video generation to 15 seconds with multi-turn conversational editing and a 30% price cut; ChatGPT's image generation was upgraded to Images 2.6, with 50% faster speed and GIF support; Anthropic launched read-only personal finance features for U.S. paid users, emphasizing privacy and security; Baidu open-sourced Early 5.5 FlyerFlashbase, with 28 billion parameters and 1.4 billion activated, suitable for research fine-tuning; and DeepSeek Harness reached 1.54 million daily active users, with a rapidly developing plugin ecosystem and plans for an official marketplace. Overall, AI technology is accelerating its penetration into various industries.