- NATURAL 20
- Posts
- GPT-6 Cyber Finds New Security Flaws as Google Expands AI Video
GPT-6 Cyber Finds New Security Flaws as Google Expands AI Video
PLUS: Anthropic's AI research may be accelerating, Perplexity launches Teammates, and Meta acquires Conductor.
Wes Roth
November 06, 2026
The AEO work, without the AEO hire.
Most AI visibility tools show you where you rank, then leave the work to you.
ChatFeatured is an AI search agent that does it: finds the questions competitors win, drafts the content in your voice, and tracks the lift.
Your team decides what ships. Month to month.
Today:
GPT-6 Cyber helps OpenAI researchers uncover 81 new Firefox vulnerabilities
Gemini Omni 1.2 Flash generates 15-second videos with more precise edits
Claude Personal Finance connects accounts with read-only access
Perplexity launches proactive AI Teammates
METR estimates faster AI-assisted research at Anthropic
AI Moves Into Security Research, Video Production and Personal Finance
AI developers are putting more specialized systems into real workflows. OpenAI has introduced a powerful cybersecurity model, Google is expanding AI video generation, and Anthropic is connecting Claude to the financial information users choose to share.
The week also brings more proactive agents and tools for AI-assisted software development. As these capabilities spread, access controls, transparent limits and the ability to check results become increasingly important.
OpenAI introduced GPT-6 Cyber, a model designed to help approved security researchers discover vulnerabilities, develop proof-of-concept exploits in authorized environments and verify software fixes. The company says it improves on GPT-5.6 Cyber while more reliably staying within the task boundaries set by researchers.
GPT-6 Cyber scored 48.5% on ExploitGym, compared with 32.1% for GPT-5.6 Cyber. That is a 16.4-percentage-point improvement, or 51% relative growth. On a separate web penetration-testing evaluation, it identified 40.3% of vulnerabilities compared with 32.0% for GPT-6 Astra.
OpenAI says its team used the model to develop 948 privilege-escalation exploits in Linux and identify 81 previously unknown high-severity vulnerabilities in Firefox. It also assisted with candidate fixes, and Mozilla has released patches for many of the reported Firefox issues.
The model was also evaluated on whether it respected testing boundaries. In ExploitGym Honeypot, it avoided exploiting out-of-scope services in 99.93% of cases without additional deployment prompts and 100% with them. These are controlled evaluation results rather than a guarantee of behavior in every environment.
OpenAI classifies GPT-6 Cyber as Critical in cybersecurity under its Preparedness Framework. Initial API and Codex access is restricted to a small group of Daybreak Red organizations that have received separate approval for this model. The company plans to expand access gradually to more verified defenders.
Google released Gemini Omni 1.2 Flash on November 5, raising the duration of native video generation and editing from 10 seconds to 15 seconds. Users can choose clip lengths between three and 15 seconds in one-second increments.
The company says the model better preserves faces, clothing, packaging details, props and visual styles as scenes and camera angles change. Creators can supply images, short motion clips and supported audio references to help guide generated content.
Its multi-turn editing feature lets users request changes to a product, background, lighting, weather or aspect ratio without rebuilding a whole clip. Google also highlights clearer lip-sync, more natural multi-speaker dialogue and smoother camera movements.
API pricing for 720p generation is $0.07 per second, which Google says is 30% lower than the previous Flash release. Teams can start with lower-resolution drafts and upscale their preferred results to 1080p or 4K.
Gemini Omni 1.2 Flash is generally available through the Gemini API and Google Cloud business tools, while integrations into Flow, Google Vids and other products vary by plan or rollout. Audio-reference features require allowlisting, and Google notes remaining challenges with rapid motion and complex physics.
Anthropic introduced Personal Finance, a dedicated Claude experience for understanding bank and credit-card activity through natural-language questions. It is beginning to roll out in beta to Pro, Max and Ultra users in the United States.
People can ask Claude to group expenses, compare month-to-month spending, examine recurring charges or understand upcoming bills. The assistant can also create editable charts, calendars and dashboards from the records users authorize it to access.
Connections are handled through Plaid, where users authenticate directly with their bank. Claude receives read-only data for selected accounts, including balances, credit limits and up to two years of transaction history, but it does not receive bank passwords or gain authority to transfer money or place trades.
Anthropic says conversations in the dedicated Personal Finance hub are not used for training its models. Users can unlink an account whenever they choose, and the company says the feature does not display sponsored recommendations or promote its own financial products.
The experience is meant to help people understand financial information, not provide investment advice or replace regulated professionals. Anthropic advises users to check important calculations against their bank records.
🧠RESEARCH
MoviEdit tackles long-form video editing by preserving changes across different shots, camera angles, and subject appearances. Its researchers built 87,000 paired editing videos and a new evaluation benchmark. They report better results than existing open-source systems and competitive performance against leading commercial models for text-guided and image-guided edits in experiments.
Kling researchers examine how multimodal understanding shapes video generation through captioning, prompt enhancement, reference grounding, and reward models. The report describes a shared model spanning video, images, audio, and mixed inputs. Its main lesson is that consistent video depends on accurately interpreting instructions, evaluating outputs, and systematically tracing generation failures.
LoCo teaches reasoning models when to stop rather than simply shortening every answer. It identifies redundant reasoning, creates verified shorter continuations, and learns from preference pairs. Across mathematics and code benchmarks, researchers report reductions of 41.7–45.4% and 23.4–36.3% in completion length respectively, while maintaining or improving accuracy on tested tasks.
📲SOCIAL MEDIA
🗞️MORE NEWS
Perplexity introduced Teammates, agents that use a user's Computer projects, tools, skills and memory to take on continuing work. They can collaborate with one another, contact coworkers in Slack and respond to changes without needing a new prompt each time. The feature launched on the web for Max subscribers, subject to existing project permissions.
Independent evaluator METR estimates that Anthropic's model-capability progress is now around twice the earlier rate, with an internal benchmark trend showing a 1.9-fold acceleration since October 2025. METR believes AI assistance is likely responsible for much of that shift, but emphasizes substantial uncertainty and says Anthropic redacted significant supporting evidence. It does not conclude that autonomous AI already performs most frontier AI research.
Tencent launched Hy4, a roughly 770-billion-parameter model adding multimodal understanding and stronger long-running agent workflows. It can analyze images and charts and help with coding and computer interaction; Tencent says it restricted cybersecurity training toward defensive use. API access was announced, while downloadable weights and additional safety materials were described as forthcoming.
Moonshot released the weights for Kimi K3.1 Pro following its October model introduction. Its mixture-of-experts architecture has 2.8 trillion total parameters, approximately 104 billion active per token and a one-million-token context window, with about 1.56 TB of files on Hugging Face. Self-hosting requires significant hardware and use is governed by Moonshot's custom license.
Cognition released Devin Review 2.0, updating its AI-assisted pull-request review tool. It says the new version is up to five times cheaper and approximately 1.5 times faster, aiming to make review more practical across everyday development work. Those improvements are company-reported and will depend on the review workload.
Conductor announced that it is joining Meta, bringing its parallel coding-agent workflow technology into a larger AI organization. New Conductor sign-ups are closed; Conductor Cloud becomes read-only November 13 and shuts down November 20. Local Conductor will continue receiving model updates and critical fixes until early 2027.
OpenAI announced plans to acquire Pierre Computer Company, whose Code Storage and open-source code viewing tools already contribute to the Codex experience. The company wants to improve support for branches, repositories and reviews as coding agents generate more software, alongside its Pando source-control backend. The announcement describes a planned transaction, not a completed acquisition.
What'd you think of today's edition? |
Reply