OpenAI Releases GPT-6 Cyber for Security Research as Perplexity and Hugging Face Push Agent Tooling

Key Takeaways

1. OpenAI released GPT-6 Cyber, a security-research model it says improves both capability and alignment, with access limited to some Daybreak Red customers and a new Daybreak Red program for critical-infrastructure defenders.
2. Hugging Face open-sourced an ML Intern Slack bot and published a free MoE training handbook backed by 800,000 GPU hours, arguing 18 of the current top 20 open models are MoE.
3. Perplexity launched Perplexity Teammates, a work-oriented personal agent, reporting about 10,000 teammates in its internal workspace over 30 days.
4. Google's Gemini Omni Flash 1.2 extends video generation and editing to 15 seconds and cuts API pricing 30% versus 1.1, with Runway also adding the model.
5. Tesla plans to phase out Cybertruck production around mid-next-year in Texas to free capacity for Cybercab, while SpaceX targets recovering both Super Heavy and Starship on an upcoming flight.
6. A new DeepFinance benchmark targets long-horizon high-end finance work with 21 tasks estimated at 20-40 human expert hours each.

1. AI Models and Security Research

1. OpenAI released GPT-6 Cyber, saying it raises both security-research capability and model alignment; the model underpins previously announced Linux kernel research and is already used by Mozilla, Palo Alto Networks, Calif, Cisco and OpenAI's own security teams, with access currently limited to some Daybreak Red customers and set to expand. Greg Brockman called it a major advance for security research and said OpenAI invested heavily in training the model to respect operator-specified scope, while also announcing Daybreak Red for approved critical-infrastructure defenders to conduct authorized testing. A cited post said the model found 948 vulnerabilities in Linux and 111 renderer RCEs in Firefox, including 81 new discoveries and 30 exploits of known-but-deemed-unexploitable flaws. 1 2 3
2. A forwarded account of testing pre-release GPT-6 Astra described a SymPy-based math grader bypass in which lookalike characters caused the grader to execute a shell command printing “4242” while still returning the correct answer and marking it correct; the post argued such flaws are hard for humans to find even with fuzzing because they require reasoning across input validation and evaluation, and said the goal is to expand Daybreak to organizations of all sizes, independent researchers and open-source maintainers. 1
3. Hugging Face said 18 of the current top 20 open models use mixture-of-experts architecture and that the dense-model era is largely over, but MoE training knowledge remains concentrated in a few teams; it invested 800,000 GPU hours in experiments and released a free, open “The Ultimate MoE Training Handbook” of about 70,000 words with 55 interactive charts and runnable code covering expert count and size, shared experts, all-to-all communication overhead, router choices, auxiliary losses and upgrading dense checkpoints. It also launched an open-source ML Intern Slack bot for evaluating models, building datasets and fine-tuning models in a workspace. 1 2
4. Google's Gemini Omni Flash 1.2 extends video generation and editing to 15 seconds, adds deeper cinematic camera control and multi-turn conversational editing, reduces visual drift and lowers API pricing 30% versus Omni Flash 1.1; 720p video is priced at $0.07 per second, with access through Flow, Google Vids, the Gemini app and the API. Runway added the model, supporting creation and editing of videos up to 15 seconds and 4K upgrades, and also added GPT Image 2.6 with a claimed 50% speed increase, GIF and SVG export, and precise edits that preserve subject and brand details. 1 2 3 4

2. Agents and Developer Tooling

1. Perplexity launched Perplexity Teammates, a work-oriented personal agent that shifts interaction from writing prompts in a session to giving high-level goals; it opened today to web Max users as a research preview, with mobile and desktop native apps to follow. The agent can proactively contact users, colleagues and other agents in the workspace and close work loops without repeated guidance, with Perplexity citing deep research, data analysis, QA, ticket triage and code review as strengths. Internally, Perplexity reported about 10,000 teammates in its workspace over 30 days, 20 times its headcount, with about 1,000 active the previous day; median agent hours per employee rose 7x since September and one-fifth of employees exceeded 40 hours per week. 1 2 3
2. Aravind Srinivas said agent prompts are increasingly sent by teammates to each other rather than typed by humans, and that agents connect to many cloud tools and process context beyond human tracking capacity, while difficult pre-release decisions, restraint and judgment remain with people; he argued agents will increasingly be used by other agents because orchestration is simpler and can be centralized around git as a compounding system of record. 1
3. Hugging Face said it is livestreaming training and, per a cited post, is using reinforcement learning to train LoRA adapters for Liquid AI's LFM2.5/LFM3 and thewebAI's Twil-9B to improve efficiency and accuracy on agent tasks. 1
4. A forwarded tool description for Raindrop AI said it offers PR-level agent testing with one-click tests against real scenarios on every PR, handling sandboxing, fixtures, real and synthetic data and evaluations. 1

3. Benchmarks, Enterprise AI and Infrastructure

1. DeepFinance was released as what its backers call the first ultra-long-horizon benchmark for high-end finance work, with 21 tasks built on real trading data by experts from Evercore, JPM, UBS and HBS/Wharton backgrounds, each estimated at 20-40 hours for a human professional. The release argued existing finance benchmarks are limited to document QA or simple numerical correctness, while real high-end finance work involves assumptions and conflicting evidence, and that DeepFinance aims to evaluate agents end to end on workflows such as due diligence for a deal above $100 million. 1 2
2. OpenAI and JPMorgan Chase announced a strategic partnership to explore how advanced AI can support its corporate business, including helping engineers build and security teams defend; JPMorgan will use Codex and OpenAI's cybersecurity model and deploy models company-wide for enterprise transformation and knowledge work. A cited post said OpenAI's revenue mix has shifted from about 30% enterprise and 70% consumer to an enterprise majority, with consumer still growing. 1
3. Elon Musk said Tesla will phase out Cybertruck production around mid-next-year at Gigafactory Texas to free space for higher Cybercab output, while a forwarded post called Cybertruck an important technology development platform whose spirit continues in Cybercab. He also said SpaceX will attempt to recover both Super Heavy and Starship on an upcoming flight, which if successful would be the first complete return of all orbital rocket components, and a cited post said the flight would be the first orbital Starship launch carrying advanced Starlink satellites. 1 2 3 4
4. Musk forwarded details of a planned SpaceX data center in Cochise, Arizona, described as about 6 GW across nine buildings, using no industrial water and with no gas turbines or on-site generation, about three miles from the nearest homes; the forwarded post listed local power discounts, annual funding, taxes, investment, jobs and environmental commitments. 1 2

4. Industry Views and Adoption Signals

1. Ethan Mollick argued a potential financial downturn in AI stocks would not substantially change AI's trajectory: AI will not disappear and enterprise adoption and development will not slow, though policy and effort are still needed to ensure good use. He also criticized AI labs, policymakers and NGOs for underinvesting in near-term “national AI public-good moonshots” and suggested universal tutoring, medical second opinions, financial literacy and fraud prevention, and easier access to government services, saying his own 20 years building tutoring and financial-literacy programs were technology-limited and that current models could produce real gains with careful design and research, though failure is possible. 1 2
2. Kai-Fu Lee said AI is already smart enough that CEOs should not limit it to shallow tasks such as customer service or document checking, and discussed shallow enterprise AI use, open versus closed models and how leaders should drive AI transformation; a reader described his book AI Native as more of a practical manual, arguing CEOs should learn to use AI tools themselves rather than delegate adoption to siloed departments. 1 2 3
3. Swyx argued evaluators cannot trust labs not to optimize for benchmarks intentionally or unintentionally, and called for credible third parties to rebuild entire evaluations in the same spirit, from shuffling multiple-choice options to designing new questions that look different but test the same reasoning steps; a cited post said Kimi K3.1 “recalled” instructions never given on Terminal Bench that matched hidden tests, and failed when the task was changed to contradict them. 1
4. At AI Engineer Shanghai, Ryo Lu's talk argued tools should help people form their own answers rather than skip thinking, and he said he moved from San Francisco back to Taipei to preserve freedom to make different choices; attendees described the talk as a highlight and stressed standards of quality and beauty as a human distinction from AI's speed and quantification. Another attendee relayed a view that teams of 2-3 people are enough and current startup teams should not hire, that US AI social is crowded and moving offline, and that very vertical niches are the remaining path. 1 2 3 4