GPT-6 Cyber arrives and Gemini video reaches 15 seconds
Security teams have a new model for validating vulnerabilities, and Gemini video creators can keep editing the same clip. Cursor adds app hosting, while GitHub moves merge readiness to the top of pull requests.
01
Model updates
3 stories
2026-11-05OpenAI
OpenAI releases GPT-6 Cyber with vetted Daybreak Red access
OpenAI introduced GPT-6 Cyber for authorized vulnerability discovery, validation, and patching. Vetted security teams can access the model through Daybreak Red, with identity checks and organizational access and use restrictions.
On ExploitGym, which asks models to exploit a specified vulnerability, it scored 48.5% versus 32.1% for GPT-5.6 Cyber. A separate isolated test included an exploitable service outside the task's authorized scope. Without deployment safety prompts, the new model exploited it in 0.07% of graded runs, compared with 5.51% for the previous Cyber model.
2026-11-05Thinking Machines Lab
Inkling 2 Small comes to Tinker, with weights planned for next week
Thinking Machines Lab introduced Inkling 2 Small, a mixture-of-experts model with 594 billion total parameters and 17 billion active per step. It accepts images and supports adjustable reasoning effort. The company reports a score of 31 on the Artificial Analysis Intelligence Index.
The first release is available through Tinker with closed weights. The lab plans to publish the full weights under Apache 2.0 the following week, after external safety testing and red teaming are complete.
2026-11-05Baidu
Baidu releases pretrained ERNIE 5.5 Flash weights
Baidu released the pretrained weights for ERNIE 5.5 Flash Base. The model card lists 280 billion total parameters, of which 14 billion activate per token. Pretraining used 33.5 trillion tokens and contexts of up to 128K.
The checkpoint comes from the end of pretraining and gives teams a starting point for domain adaptation, instruction tuning, and reinforcement learning. Fine-tuning and inference support up to one million tokens of context. Baidu offers FP8 and BF16 weights under a dedicated model license.
02
Products and developer tools
6 stories
2026-11-05Google
Gemini Omni 1.2 Flash extends native video to 15 seconds
Google made Gemini Omni 1.2 Flash generally available in the Gemini API and Gemini for business. The native generation and editing limit rises from 10 to 15 seconds, with one-second increments from 3 to 15 seconds. At 720p, output costs $0.07 per second, 30% less than Omni 1.1.
Users can revise the same clip conversationally to change its background, lighting, or aspect ratio. They can also guide generation with images, motion references up to three seconds long, and audio. The workflow supports drafting at 360p or 720p and upscaling selected clips to 1080p or 4K.
2026-11-05Cursor
Cursor cloud agents can deploy apps to shareable links
A Cursor cloud agent can scaffold a web app from a description, provision Postgres as needed, configure access through Cursor single sign-on, and prepare hosting. The user chooses a domain and publishes the app to get a shareable link.
The feature supports new or existing repositories on Cursor Origin and GitHub, with GitLab and other providers planned. Pre-publication checks look for problems such as credentials embedded in files. Once the checks pass, the user still has to publish explicitly.
2026-11-05Anthropic
Claude starts a US beta for personal finance queries
Claude's Personal Finance beta starts rolling out this week to US Pro, Max, and Ultra users. Users can authorize a connection to bank and card accounts through Plaid and ask about spending categories, upcoming bills, or duplicate charges. Access is read-only and the feature does not provide investment advice.
Claude can read balances and up to two years of transactions for the accounts the user connects. Anthropic says conversations in the Personal Finance hub are excluded from model training. Information users share outside the hub remains subject to their usual privacy settings.
2026-11-05Meta
Meta rolls out live voice and expressive avatars in Muse
Meta began rolling out live voice and expressive avatars in the Muse app on mobile, web, and Mac in the US. The phone icon opens a conversation. Users can interrupt a reply and see the avatar react with expressions and gestures as they speak.
The interface offers voice only, an avatar alongside the conversation, or a full-screen avatar. Meta previewed live conversations at Connect and is now making them available in the app. Voice conversations on AI glasses and Muse Charm are still planned.
2026-11-05AWS
AWS shows how an agent can get a cited web answer in one call
AWS published an integration walkthrough for Web Answer on Bedrock AgentCore. An agent sends a natural-language question through Gateway. The AWS-managed tool plans parallel searches, reads pages, searches again if evidence is thin, and returns an answer with source links as an MCP tool result.
AWS also distinguishes Web Search from Web Answer. Web Search returns ranked URLs, titles, and snippets for the application to process. Web Answer runs the retrieval and synthesis itself, making it an option for a focused fact-finding task within an agent workflow.
2026-11-05GitHub
GitHub brings merge readiness to the top of pull requests
GitHub opened an opt-in preview of the new pull request Overview, with rollout continuing over the weekend. Merge readiness and the next available actions move to the top, where reviewers can find failing checks, unresolved comments, and missing approvals.
The preview also lets users filter for unresolved threads or their own comments, or sort newest first. Users can opt in from the header, where actions such as marking a draft ready or closing the pull request appear when permissions allow.
03
Research and evaluations
3 stories
2026-11-05METR
METR reviews Anthropic's early warning assessment for automated R&D
METR published its external review of the automated R&D section of Anthropic's August risk report. It agreed that models deployed internally as of July 15 posed low risk of directly automating key R&D or rapidly causing catastrophic harm through accelerated capabilities progress.
The difference concerns an earlier warning about whether AI automation might substantially speed up AI R&D. The METR researchers on the review and relevant Anthropic staff both lean toward the threshold not having been crossed. METR considers it plausible that the threshold had already been crossed and says the public report understates the acceleration.
2026-11-05Anthropic
Anthropic examines 207,000 research sessions
Anthropic used privacy-preserving methods to study 207,049 research sessions on consumer plans from August 2 to September 30. Claude Chat and Cowork accounted for 128,914 sessions, and Claude Code for 78,135. The study classified the sessions by field and task.
Data analysis accounted for about one-third of the Chat and Cowork sample and nearly half of the Claude Code sample. Idea generation and data collection appeared less often, even though researchers ranked those two activities highest for scientific discovery in an accompanying survey.
2026-11-05Reflection
Reflection details how it trained and evaluated Beam for safety
Reflection published the safety training methods for its open-weight model Beam and shared its Iris Safety Evaluations. In Iris conversation tests, Beam refused 89% of harmful requests and complied with 82% of harmless requests involving safety topics. The tests span 1,500 human-written conversations across 15 harm categories.
Reflection trained separate teacher models for task capability and safety behavior, then distilled both into Beam. Because released weights can be modified, the team separately assessed capabilities that could be misused and the deployed model's willingness to refuse requests.
04
Industry and security updates
2 stories
2026-11-04Conductor
Conductor joins Meta and sets dates for its cloud shutdown
Coding tool Conductor announced that it is joining Meta. Conductor Cloud becomes read-only on November 13 and shuts down on November 20. The local Mac app will keep receiving new models and critical fixes until early 2027.
Users can transfer cloud workspaces and their conversations to a Mac or export all their data from account settings. The transfer command is available in a cloud workspace's context menu.
2026-11-05GitHub
GitHub secret scanning adds Starlink API credential detection
GitHub added detection for Starlink public API v2 secrets, and SpaceX joined its secret scanning partner program. When GitHub finds a corresponding partner secret in a public repository, it notifies the provider so the credential can be revoked or rotated.
The update also adds false-positive filtering to detectors for Alibaba, AWS, Google, OpenVSX, Pydantic Services, and Tencent.