This round of updates focuses on handling multiple types of data, managing content and controlling interactions with AI. Cloudflare is expanding the Clef family and Workers infrastructure, while ElevenLabs is adding tools for creative work and calls. Announcements from OpenAI also highlight applications of agents in security, workflow optimization and the detection of generated text.
Cloudflare launches Clef-omni, expanding multimodal inputs
Cloudflare announced Clef-omni, a model that accepts audio, video, text and image inputs, and released open weights on HuggingFace. The model supports wav or mp3 audio and mp4 or webm video. Clef-omni is built on the Qwen3-Omni-30B-A3B-Instruct mixture-of-experts architecture, using its input-processing core while removing the text-to-speech output components. In the same update, Cloudflare said it had made Clef faster and reduced Clef-flash pricing to below Jev's. However, the announcement provided no measurements of performance improvements or specific prices. The notable change is the ability to feed audio and video into a single model in the Clef family.
Source: Cloudflare
ElevenCreative brings brand identity into content creation workflows
ElevenLabs introduced Brand Kit, now available in ElevenCreative, to store colors, fonts, logos, reference images and brand rules as a shared object in a workspace. Users type @ to attach a kit to a prompt in Image & Video, Flows or the ElevenCreative agent; its assets and rules guide the generated content. The same kit can be used across tools and by different team members, reducing the need to describe the brand identity again at each step. Agencies or teams managing multiple clients can create separate kits and switch between them. ElevenLabs is also developing support for branded voices and audio, but this has not yet been announced as an available feature.
Source: ElevenLabs
Sophos reports shorter investigation times with OpenAI models
Sophos combines OpenAI models through Daybreak with its own threat intelligence, response workflows and security expertise. According to the announcement on OpenAI's website, investigation time fell by 96%; average response time for cases using agents dropped from about 38 minutes to 89 seconds. The source also states that AI handles 52% of MDR cases from end to end. Agents gather customer context, detections and indicators of compromise, then follow a plan–execute–review loop. A summary and proposed response are passed to an analyst for review; potentially destructive actions still require appropriate human oversight.
Source: OpenAI
Samsung wins six awards at Better Choice Awards 2026
Samsung was honored in six categories at the Better Choice Awards Gala on October 4, 2026. The awards are held under the direction of the Ministry of Science and Technology and jointly organized by VCCorp and the National Innovation Center. Galaxy Z Fold8 won the “Trendsetting Technology Product” award with 33,931 reader votes; Galaxy S26 Ultra won Better List's “Favorite Innovative Breakthrough Smartphone” category. Galaxy Watch9 and Galaxy Watch Ultra2 jointly received an award for AI in proactive healthcare. The remaining recognition went to Galaxy AI's Vietnamese-language AI experience, the SmartThings ecosystem and Samsung Vietnam as one of the “Top 5 Outstanding FDI Partners of the Year,” spanning devices, software and its role as a partner.
Source: Samsung
Cloudflare adds CPU and memory profiling for Workers
Cloudflare announced CPU and memory profiling tools for Workers and Durable Objects. From Workers Observability, users can request on-demand profiling of a live Worker, view an interactive flamegraph and download the results file for further analysis. The tools can be used through the CLI or Cloudflare Dashboard. In the dashboard, users select a duration and Worker version; that version needs enough traffic for profiling to succeed. Each rectangle in the flamegraph represents a function call, with its width reflecting CPU time or memory usage. Cloudflare said its internal teams have used the tools to identify optimization opportunities, fix out-of-memory errors and improve performance.
Source: Cloudflare
ElevenAgents adds synthetic voice detection for calls
ElevenLabs announced a feature in ElevenAgents that analyzes the first few seconds of a call to classify the voice as human or AI-generated. Businesses set their own rules for what happens next, such as requiring authentication, routing AI callers to an agent with specific restrictions or blocking sensitive requests. Detection works on live audio and does not depend on watermarks. At the time of the announcement, the feature was available to enterprise customers supported by the Forward Deployed Engineering team; broader availability was planned for the end of the month, according to the announcement. ElevenLabs is also joining the Personal Agent Protocol as a design partner. The group, led by Meta and Sierra, is developing ways for personal agents to identify themselves, specify whom they represent and define the scope of their actions.
Source: ElevenLabs
Asana tests history optimization for browser agents
According to OpenAI's article, Asana uses GPT‑6 Astra in Codex to test and optimize browser agent workflows powered by GPT‑6.1 Sol. The study included 144 runs using GPT‑6.1 Sol and three models referred to as Model A, B and C. StackAI enables workflows for navigating websites, filling out forms and gathering information without writing code. The inefficiency identified was that the history of page text and screenshots was not cached, causing every request to resend the full history at full cost. The most effective policy accumulated up to 20 images before retaining only the latest one, combined with a larger history budget. All configurations collected the same six information fields for 32 books from a public demo catalog.
Source: OpenAI
The Deno team joins Cloudflare to simplify self-hosting
The Deno team is joining Cloudflare to simplify self-hosting Workers and Durable Objects, helping developers use these components in more environments. Ryan Dahl said the team is unifying workerd and celld. Although workerd is already open source, its support for Durable Objects is currently limited to a single running instance. The stated direction is a distributed version in which the platform handles placement, routing and durable storage across multiple machines. celld is intended to become a binary written in Rust, with an object storage service as its only external dependency. This is a direction for reducing operational complexity, not confirmation that the distributed version is complete.
Source: Cloudflare
Scribe generates word-level timestamps during transcription
Scribe calculates word-level timestamps during transcription with a single API call. Its structured output consists of tokens classified as words, spaces or audio events; the last category can include laughter, applause and background sounds. Conventional transcription requires an additional step to align text with audio and determine timing, increasing processing time and computational demands. Scribe supports up to five channels, transcribed independently. Timestamped data and audio classifications can support synchronized subtitles, searchable transcript archives and automatic clip creation for social media.
Source: ElevenLabs
OpenAI outlines text watermarking and detection limits
OpenAI describes text watermarking as a way to meet the EU AI Act's requirement to identify generated content in a machine-readable format. API customers worldwide can enable the option for certain models; it is off by default. The company plans to add invisible watermarks to eligible text output from ChatGPT and Codex in the EU in the following weeks. The textGrain technology embeds statistical signals in word choices for a detector to check. Access to the detector is initially limited to approved researchers and specialist organizations. OpenAI plans to open-source the technology, while noting that detection works better on longer passages, is significantly weaker for mathematics and can drop sharply when words are replaced with synonyms.
Source: OpenAI
Quick summary
Cloudflare is expanding Clef's multimodal inputs, adding resource profiling tools and working toward simpler self-hosted infrastructure.
ElevenLabs is focusing on shared brand guidelines and rules for handling callers using synthetic voices.
The Sophos and Asana cases illustrate the role of AI agents in security investigations and browser workflow optimization, with results tied to their respective announcements.
Scribe's timestamps help make use of audio recordings, while OpenAI's text watermarks support content identification but have limitations when the text is edited.
Further reading
News Roundup, October 9, 2026: More Accessible Internet Data, AI Expands in Work and Content
OpenAI rolls out text watermarking in the EU, limits access to detection
ElevenLabs reaches a $22 billion valuation, steps up conversational AI
When do AI agents need to ask permission? A perspective from OpenAI's guidance