“Stealth” model Space Bunny Alpha appears on OpenRouter and OpenCode
要闻
Space Bunny Alpha has a release date of September 23, 2026. The anonymous model has appeared on OpenRouter and OpenCode, is listed at a price of 0, and supports a 1,000,000-token context and outputs of up to 524,288 tokens.
Space Bunny Alpha, developed and operated by an anonymous third party, has a release date of September 23, 2026, and has appeared on OpenRouter and OpenCode. OpenRouter only forwards requests and is not the model’s developer, owner, or provider. The model is hosted by just 1 provider, and OpenRouter sends every request directly to that provider without selecting among providers. Its listed price is 0, with no charge for either prompt tokens or completion tokens. According to the page, the provider may retain prompts and completions but does not use them for training, while other uses are governed by the Stealth Model Terms.
According to the page, Space Bunny Alpha has a context window of 1,000,000 tokens, supports up to 524,288 tokens per completion, and allows its reasoning effort to be adjusted. It accepts text, images, and video as input and returns text. For function calling, it supports tools and tool_choice. For JSON output, it supports response_format but does not enforce a JSON schema. The page also describes the model as having coding capabilities and native support for multimodal input.
Google releases Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS
要闻
Google has launched two Gemini 3.8 TTS models supporting more than 100 languages. The Flash model scored 71.4 and ranked No. 1 on Hume AI’s Voice Design Benchmark.
Google has begun rolling out Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS through Google AI Studio, the Gemini API, Gemini Enterprise, Gemini Notebook and Google Vids. According to the company, Flash TTS expands its original set of 30 voices into an unlimited range of custom voices created with natural-language prompts, with controls for accents, emotion, line-by-line delivery, two-speaker dialogue and natural turn-taking; both models support more than 100 languages. Google says Flash TTS scored 71.4 overall and 60.8 for accent modeling on Hume AI’s Voice Design Benchmark, ranking No. 1 in both categories. Flash TTS and Flash-Lite TTS placed No. 1 and No. 2, respectively, on the Overall Quality Index. Google also says the models ranked among the leading systems in Voice Arena blind evaluations for Japanese, Brazilian Portuguese, Vietnamese, Modern Standard Arabic, Mexican Spanish and Hindi. Voice replication requires a verbal consent recording made by the voice owner and matching the reference speaker, while all audio generated by Gemini Audio models carries an embedded SynthID watermark.
OpenAI upgrades ChatGPT Voice with support for plugins and GPT-6 series models
要闻
OpenAI is rolling out an upgraded ChatGPT Voice globally through the latest app, adding support for email, calendar, Slack and other plugins, as well as GPT-6 Astra, Sol and Luna, while extending it to ChatGPT Work on web and mobile.
OpenAI announced an upgrade to ChatGPT Voice, with the related features now rolling out globally through the latest app. According to the company, users can invoke email, calendar, Slack and other plugins during voice conversations, including plugins developed by third parties, while GPT-6 Astra, Sol and Luna can also power the voice feature. ChatGPT Voice is also available in ChatGPT Work on web and mobile, allowing users to create documents, presentations, websites and spreadsheets by speaking, or handle complex tasks in a browser.
Claude Code cloud sessions officially launch, with trial credits available to subscribers
开发生态
ClaudeDevs announced that Claude Code cloud sessions have exited research preview and are now generally available, with tasks continuing after a computer is shut down. Existing Pro and Max subscribers can claim one-time credits of $100 and $250, respectively.
ClaudeDevs announced that Claude Code cloud sessions have exited research preview and are now generally available. According to its explanation, cloud tasks run on infrastructure hosted by Anthropic, allowing sessions to continue after users close their laptops or shut down their computers. Users must first connect GitHub, then start a session from the Code page on the Claude web app, the Code tab in the Claude mobile app, the desktop app, or by entering claude --cloud in the CLI. Existing Pro and Max subscribers can claim one-time credits of $100 and $250, respectively, with one claim allowed per account. The claim deadline is October 7, but the source does not specify the year. The credits apply only to cloud sessions, are calculated separately from regular usage limits, and are used automatically after a cloud session starts. Users who reach their local usage limit can continue using cloud sessions until the dedicated credits are exhausted.
Qwen releases five Qwen-Audio-3.1 models, with APIs for four now available
模型发布
Qwen has released five Qwen-Audio-3.1 models spanning audio understanding, generation, interaction, and creation. APIs for four of the models are now available on the Qwen AI platform, with ASR-Next the only exception.
Qwen has released five models in the Qwen-Audio-3.1 series, with APIs for four of them now available on the Qwen AI platform. The ASR-Next API is not yet live and, according to the company, will be offered later. The lineup consists of ASR, TTS, Realtime, ASR-Next, and TTS-Next. ASR adds transcription polishing and speaker-attributed transcription. TTS-Next can generate voices, sound effects, and ambient audio from text, timestamps, and reference audio. Realtime supports full-duplex conversations and can call tools during a conversation. Qwen said it has reduced prices across its Qwen-Audio speech models, with TTS prices down by about 70%, Realtime by about 85%, and ASR by up to 95%.
iFLYTEK releases Spark-ASR-2.0, with a gradual rollout to 讯飞输入法
模型发布
iFlytek has released the Spark-ASR-2.0 large speech recognition model and plans to roll it out gradually to iFlytek Input Method from September 24 while offering an API through iFlytek Open Platform. It supports switch-free recognition of dialects from 202 cities across China, with inference costs about 10% higher than version 1.0, according to the company.
iFlytek has released the Spark-ASR-2.0 large speech recognition model and will begin rolling it out gradually to iFlytek Input Method on September 24. The model will also be offered as an API service through iFlytek Open Platform, and its demo page is already available. It is based on the Spark-Audio-1.0-Preview speech foundation model. According to the company, the system combines non-autoregressive recognition with LLM-enhanced autoregressive recognition, joint enhancement of mixed Chinese-English text and acoustics, and dynamic context injection. These methods are intended to improve recognition of mixed Chinese and English, dialects, technical terms, high-noise audio, and low-volume audio while producing more fluent and standardized transcripts. The model supports switch-free recognition of dialects from 202 cities across China. The company says its overall inference cost is about 10% higher than Spark-ASR-1.0. Spark-ASR-2.0 will later be introduced gradually to the 讯飞 AI 眼镜, 讯飞智能办公本, and 讯飞听见 products.
Alibaba open-sources Logics-Parsing-V3 for cross-page parsing of long documents
模型发布
Alibaba has open-sourced the 0.8B-parameter Logics-Parsing-V3, extending document parsing from single pages to cross-page long documents by carrying structural state between pages; its long-document tests cover inputs of up to 50 pages.
Alibaba has open-sourced Logics-Parsing-V3 for structured long-document parsing, extending Logics-Parsing-v2 from single-page recognition to cross-page processing. The model reads pages sequentially through non-overlapping sliding windows, using --window_size 2 by default for multi-page inputs and always using 1 for single-page inputs. Each window combines the current page with a compact structural state passed from the preceding window to reconstruct heading hierarchies, merge cross-page elements, and connect visual content with text. It also supports complex layouts, scientific formulas, and chemical notation. According to the project, the 0.8B-parameter model achieved the highest Overall score among the models compared on MPDocBench (MPDocBench-Parse) and led on Truncated Text Edit and Heading TEDS.
The project also created MPDocBench-Long by concatenating complete source documents with similar types and resolutions. It contains 5 length ranges with 50 sequences in each range and a maximum length of 50 pages. According to its description, under the 1M inference setting, Logics-Parsing-V3 maintained more consistent parsing quality than Unlimited-OCR as document length increased. inference_v3.py accepts images, PDFs, and directories of page images, and produces markdown.md, hierarchy.md, raw_output.dsl, and result.json. The model is available from Hugging Face or ModelScope. Inference efficiency was tested on 420 MPDocBench documents totaling 3,135 pages, using NVIDIA H100 GPUs with batch size 1 for every model.
Fireworks launches Ember-1 alongside custom training support
模型发布
Fireworks has made Ember-1, a specialized model based on Kimi K3, available alongside custom training support. The company says it uses 40% fewer tokens while maintaining Kimi K3 quality for coding and multi-turn Agent workloads.
Fireworks released Ember-1, trained from Kimi K3, and made it available with custom training capabilities on the Fireworks Training platform. The company says Ember-1 uses 40% fewer tokens while maintaining Kimi K3 quality, reducing the cost of coding and multi-turn Agent workloads. Its team ran more than 50 training experiments and over 200 evaluations on Fireworks Serverless Training, and developed new training algorithms to shorten reasoning while preserving accuracy.
According to Fireworks, a reasoning model can spend more than 90% of its generated tokens on internal reasoning, while multi-turn interactions repeatedly reload earlier content, causing context to grow roughly quadratically with the number of turns. Fireworks says Kimi K3’s reasoning was shortened by 35–50% without reducing accuracy across 7 benchmarks and production traffic from 2 customers. Bedside Bench contains 500 clinical cases across 10 specialized categories. The cost analysis used public Kimi K3 API prices of $3/M tokens for uncached input, $0.30/M for cached input, and $15/M for output. The company says Ember-1 was on or near the quality-cost Pareto frontier for benchmarks with more than 50 samples and outperformed K3-low.
Black Forest Labs releases open-weight robotics model FLUX 3 Action
模型发布
Black Forest Labs released the open-weight robotics model FLUX 3 Action. Its single-step 7B version posted a 38.3% ± 0.38 success rate on RoboLab, and the company says it is 1.34x–2.28x faster than Pi0.5.
Black Forest Labs released the open-weight robotics model FLUX 3 Action and published a report covering its training and fine-tuning. The company says the single-step 7B checkpoint achieved a 38.3% ± 0.38 success rate on RoboLab, above Cosmos 3 Nano’s 36.8%, while the guidance-distilled checkpoint reached 42.2% ± 0.36. On workstation and datacenter GPUs, the former processed each second of robot motion 1.34x–2.28x faster than Pi0.5 running in BF16, while predicting a 2.13-second motion horizon versus 1.0 second for Pi0.5. When run in FP8, the base and guidance-distilled checkpoints were 1.52x–3.95x faster than FP8 Cosmos 3 Nano across consumer, workstation, and datacenter GPUs.
According to the company, F3A uses multimodal Self-Flow pretraining to keep its backbone at less than half the size of Cosmos 3 Nano. It also uses distillation to remove the guidance pass and reduce the number of sampling steps to one while retaining joint video and action prediction. On a B200 GPU, the guidance-distilled version in FP8 recorded a 42.24% success rate and a 0.048 real-time factor. Cosmos 3 Nano recorded 36.8% and 0.150, π0.5 in BF16 recorded 28% and 0.032, and the step-distilled version recorded 37.92% and 0.015. Running alone on an H200 GPU priced at $3 per hour, F3A required 21.08 milliseconds per second of robot motion, with each success costing $0.09 and taking less than two minutes on average, though it could not solve every individual task. Combined with GPT 6 Astra as a hybrid policy, the system completed 90% of episodes at a cost of $8.77 and eight minutes per success. The company says this was 29% cheaper and 40% faster than the best alternative configuration.
NVIDIA open-sources speaker diarization model Nemotron 3 Diarization
模型发布
NVIDIA released Nemotron 3 Diarization, an open-weight speaker diarization model with 100 million parameters that handles up to eight speakers. The company says its 14.72% DER places it first on Diarization-Bench.
NVIDIA introduced the open-weight Nemotron 3 Diarization to mark speaker activity in live and recorded conversations. According to the company, the 100-million-parameter model supports up to eight channels and ranks first on VoiceArena Diarization-Bench with a 14.72% DER. It supports overlapping speech, chunked processing, and configurable streaming latency. Channels are assigned permanently according to when each voice first appears, allowing anonymous labels to remain consistent across audio chunks without solving the speaker permutation again for every chunk. The channels do not represent real-world identities; downstream applications can associate them with specific people through meeting metadata, user profiles, or active speaker verification models.
The model accepts 16 kHz single-channel audio, generates a Mel-spectrogram with a 10 ms frame step, stacks the features by a factor of eight into 80 ms frames, and feeds them to a 31-layer Transformer encoder with RoPE. Its default output is a [T, 8] floating-point tensor in which each value represents the probability that the corresponding speaker is active at that time, and two channels can be active in the same frame. Training used public and licensed speech, including real-world multispeaker conversations licensed from David AI and simulated mixtures spanning 21 languages generated from its audio. NVIDIA says adding the David AI data reduced compound DER from 11.19% to 10.42% at both the offline-style and ultra-low-latency operating points, an absolute decrease of 0.77 percentage points. The model outputs only speaker activity and timestamps, so producing speaker-attributed text requires combining it with ASR.
Fish Audio launches a preview of Drama 3 with natural-language voice control
模型发布
Fish Audio has released a preview of its Drama 3 text-to-speech model, allowing users to set tone, pacing, and characters with natural language. The model can also switch voices within one sentence, generate multi-character scenes, and redo only one word in generated speech.
Fish Audio has released a preview of its Drama 3 text-to-speech model, extending generation controls to natural language. Users do not need to write audio tags and can instead describe the intended tone, speaking pace, and character settings in everyday language for the model to generate the speech. The preview also supports switching between different voices within the same sentence and generating scenes with multiple characters. For localized edits, users can change a single word in the generated speech without reprocessing the entire result.
Anthropic product manager Robert Bye said the Claude mobile app now supports multiple accounts, allowing users to switch between work and personal accounts without repeatedly logging out and back in; connecting multiple Gmail or Google Calendar accounts to one Claude account remains under investigation.
According to Anthropic product manager Robert Bye, Claude has released multi-account switching for its mobile app, and the feature is now available. Users who keep work and personal activities separate can switch directly between accounts in the same app instead of logging out of the current account and signing in to another each time. This update covers only multi-account support in the Claude mobile app. Bye said the team is still investigating requests to connect multiple Gmail or Google Calendar accounts to a single Claude account and ask questions across them, and that capability is not included in this release.
Google says Gemini is adding 13 app integrations, with rollout now underway
产品应用
Google announced that it has begun rolling out 13 new app connections for Gemini. Users can invoke these tools in one chat for project management, creative asset design, and workout planning through settings, an @ mention, or a direct request.
Google announced that it has added 13 app connections to Gemini and begun rolling them out gradually. According to the company, the connections span multiple categories and let users invoke relevant tools inside Gemini to manage projects, design creative assets, and plan workouts without switching between tabs. Users can connect frequently used apps in Gemini settings or enter an @ mention in a chat to bring up the corresponding tool; they can also ask Gemini directly to use a relevant app without using mention syntax.
Google Labs launched 6 Flow tools: multitrack audio, captions, thumbnails, 3D materials, collages and motion.
Google Labs has released 6 tools in Google Flow Tools that were developed with creators from architecture, sound design and digital content. According to the company, users can build custom workflows by describing what they need. Mondo Sónico generates background ambiance, foley and contextual sound effects, automatically synchronizes them across separate editable tracks, and exports individual stems. A caption tool transcribes, styles and animates multilingual captions in a single pass, while a thumbnail tool uses one image, a headline and a prompt to generate photorealistic thumbnails for social platforms. Surface creates custom textures ranging from Minimalist patterns to Art Deco tiling and maps them onto 3D walls, ceilings and floors in real time. The other 2 tools generate dynamic mixed-media collages at scale from text prompts and convert scripts into Swiss-style motion graphics to reduce traditional keyframing work.
Google makes video generation in Google Vids free for all account users
产品应用
Google has made Google Vids’ AI generation features free for Google and Google Workspace accounts, enabling users to create 1080p videos with Gemini Omni 1.1 Flash and control scene duration and transitions.
AI video generation in Google Vids is now included in the free access available to Google and Google Workspace accounts. On desktop, users can open vids.new and select “Create AI videos” to generate and edit 1080p scenes with Gemini Omni 1.1 Flash. The tool provides controls for scene duration, style, and transitions, covering the process from initial concept creation to final export. According to Google, users do not need professional editing experience. Every video clip generated with Omni 1.1 has an imperceptible SynthID digital watermark embedded in its frames, which Google says allows viewers to verify whether the content was generated by AI.
Google also announced that Gemini 3.8 Flash-Lite text-to-speech will be added to Google Vids. According to the company, the feature will convert a user’s script into a natural-sounding voiceover and support 100+ languages. Google only said the feature is coming soon and did not provide a specific launch date. Personal account users who need more AI video generation capacity can explore Google AI plans, while Workspace Business and Enterprise plans for teams include expanded generation pools and centralized administrator controls. Plan details are available through the Google Workspace Learning Center.
Anthropic says Claude discovered a new enzyme system in bacteriophages
技术与洞察
Anthropic announced that Claude discovered a new enzyme system, ART, after roughly 950 Agents searched DNA data for 21 hours using 210 million tokens. Its function remains unknown, and the findings have been released as a pre-print.
Anthropic announced that Claude discovered a previously uncharacterized enzyme system, array-associated reverse transcriptases (ART), in one of the first projects launched after the company formed its life sciences research team in spring 2026. According to the company, the researchers only supplied the initial prompt to search for new RT examples and handled the laboratory work. Roughly 950 Claude Agents used 210 million tokens to search a DNA database for 21 hours, independently analyzing different RT families and candidates. One Agent noticed repeated DNA sequences next to an unusual-looking reverse transcriptase (RT) gene. Following additional analysis and experiments, the team identified it as ART found in bacteriophages and released a pre-print.
The RT underlying ART comes from a jumbo phage, and the enzyme itself had already been identified in previous research. Anthropic said Claude appears to be the first to connect it with two defining features: a neighboring array of non-coding DNA sequences and an accessory protein of unknown function. ART’s primary function remains undetermined. According to the company, this combination of features has previously appeared in only a handful of other systems, all of which are programmable and can perform operations such as cutting, copying, or pasting DNA. Anthropic’s Bay Area laboratory conducts only BSL-1 and BSL-2 research, does not handle pathogens capable of infecting humans, and has human scientists perform all laboratory work.
Anthropic details a Claude-assisted performance optimization method, with over 3,000 changes merged in two weeks
技术与洞察
Anthropic said it used Claude to make the core experience of claude.ai and its desktop app about 3x faster in two weeks while merging more than 3,000 changes. According to the company, the work caused no customer-facing incident or rollback.
Anthropic used Claude during a two-week sprint to optimize claude.ai and the Claude desktop app, making the core user experience about 3x faster. According to the company, more than 3,000 changes were merged without a customer-facing incident or rollback. The team focused on four journeys covering 95% of user activity. At the 75th percentile, the time from a fresh claude.ai load to a typeable page fell from 3.1 seconds to 0.55 seconds, starting a new Claude Code session fell from 0.8 seconds to 0.3 seconds, and loading a Claude Cowork cloud session fell from 2.6 seconds to 0.73 seconds. Anthropic estimates that these changes eliminate tens of thousands of user-hours of waiting each day.
The work used the beta version of Claude Tag, whose internal research model was, according to Anthropic, roughly comparable to Opus 5.5. Claude analyzed data through the Datadog MCP server, identified four journeys—launching the app, starting a new conversation, loading an existing conversation, and sending a message—and divided them into 13 measurements. The team standardized measurement from the user interaction through rendering of the result while separating client and server work. It initially listed about 20 projects and had met 12 of the 13 targets by day three. Changes included embedding a static composer in HTML, precompiling a V8 code cache, prefetching sessions on hover, and reducing sidebar re-renders by 90%. The team also created benchmarks using instruction counts, React commits, DOM mutations, and other metrics; only metrics shown to correspond to wall-clock latency were retained and added to CI.
Anthropic employee discusses Opus 5.5 writing adjustments: balancing model understanding with human readability
技术与洞察
Anthropic employee Jackson Kernion, who worked on Claude fine-tuning, said the released Opus 5.5 is the company’s first model specifically improved to address sentence clarity and information density, with a better balance between model understanding and human readability.
Anthropic employee Jackson Kernion, who worked on Claude fine-tuning, said the released Opus 5.5 received targeted adjustments to its writing and is Anthropic’s first model designed specifically to address sentence clarity and excessive information density. The Claude account said the version expresses itself more naturally, presents the most important information first, follows user-defined writing rules more effectively, and makes extended conversations easier to read. Kernion explained that as models receive more training in mathematics, code, and explaining technical issues to other large language models, the training process must also reward concise explanations that humans can understand in order to maintain balance. In his assessment, Opus 5.5 achieves a better balance between model understanding and human readability, although its writing still has room for further improvement.
DeepSeek reveals technical details of DSec, an Agent training sandbox platform
技术与洞察
DeepSeek has published a paper detailing DSec, a sandbox platform for Agent training and evaluation. A production unit with about 160 nodes serves roughly 3 million sandboxes per day, supports more than 380,000 concurrent sandboxes, and creates over 5,000 per second.
DeepSeek has published a paper describing how DSec, its sandbox platform for large-scale Agent training and evaluation, operates. According to the company, a production unit with about 160 nodes serves roughly 3 million sandboxes per day, supports more than 380,000 sandboxes running concurrently, and can create over 5,000 sandboxes per second. Through a unified SDK, DSec manages four backend types: function calls, containers, micro virtual machines, and full virtual machines. It also handles cluster scheduling, environment composition, and on-demand image loading. Agent tasks can continuously execute commands, invoke tools, and preserve state in isolated environments. The company says the platform separates sandbox execution from preemptible GPU training jobs and restricts file and network access.
罗福莉 unveils HySparse2, the core of MiMo-V3's new architecture
技术与洞察
Luo Fuli unveiled HySparse2 for planned use in MiMo-V3; according to her account, it delivers a 5.02-fold reduction in prefill FLOPs and a 4.5-fold reduction in KV cache size at one million tokens while improving four evaluation metrics.
Luo Fuli announced HySparse2, the core of a new architecture planned for MiMo-V3, and released an architecture overview and paper. She said short operations in Agentic reasoning can return long observations, while continuously expanding context increases the burden of prefill, KV cache, and retrieval. According to her explanation, HySparse2 uses two-level KV sharing to reduce computation and cache usage, while also changing token selection and the handling of recent context. Her reported results show that, compared with MiMo-V2.6’s Hybrid SWA architecture, HySparse2 delivers a 5.02-fold reduction in prefill FLOPs and a 4.5-fold reduction in KV cache size at one million tokens; MRCRv2 and RULER-v2 scores increase, while AgentPPL and LongPPL decrease.
OpenAI launches MentalHealthBench, an open benchmark for evaluating mental health conversations
技术与洞察
OpenAI has released MentalHealthBench, an open benchmark that uses synthetic conversations to test AI responses in mental health scenarios ranging from everyday emotional support to urgent crises. It was developed by more than 80 licensed experts from 22 countries.
OpenAI has launched and made MentalHealthBench publicly available to evaluate how models respond to mental health conversations in real-world situations. The tests use synthetic conversations covering everyday emotional support and crisis scenarios requiring urgent assistance. They assess model outputs for safety, contextual understanding, respect for user autonomy, and the ability to offer actionable suggestions when appropriate. The benchmark was jointly developed by more than 80 licensed mental health experts from 22 countries. Researchers can review its methodology, run the evaluations themselves, and use it as a basis for further research. OpenAI says the work evaluates model responses and does not constitute the launch of a mental health service for ChatGPT users; ChatGPT is also not a substitute for therapy or professional care.
Anthropic details the use of Claude in responding to the Ebola outbreak in the Democratic Republic of the Congo
行业动态
Anthropic said Claude is being used in the response to the Bundibugyo ebolavirus outbreak in the DRC, cutting WHO AFRO’s sitrep preparation time from a full day to under one hour while also supporting vaccine data organization and viral genomic analysis.
Anthropic disclosed that its Beneficial Deployments and Applied AI teams are working with partners convened by CEPI, including WHO AFRO and INRB, to use Claude for data processing, modeling, research evidence review, and genomic analysis in the Bundibugyo ebolavirus outbreak that has continued in the DRC since May. According to Anthropic, the outbreak has nearly 8,000 confirmed cases, and close to half of those patients have died. Community health workers record symptoms, contact histories, recoveries, and deaths through door-to-door visits, while health facilities aggregate the information through channels including WhatsApp. A Claude skill developed by WHO AFRO staff extracts case and laboratory figures from each health zone’s PowerPoint, checks them against the previous day’s report, flags changes in trends, and produces summaries. This has reduced sitrep preparation time from a full day to under one hour.
According to Anthropic, WHO AFRO’s data team also uses Claude to run multiple disease models simultaneously and produce forecasts that logistics staff can use when deciding where to locate treatment centers. No vaccine is currently approved for BDBV, while Ervebo is approved for the Zaire strain. After BDBV emerged, CEPI solicited projects involving vaccine candidates and related scientific work, and it built a dashboard with Claude to track tasks required to advance vaccine development. Claude organizes multifactor data, while experts remain responsible for selecting cohorts of serological samples and assessing Ervebo’s potential cross-reactivity against BDBV. The INRB laboratory produced nearly all BDBV genomes sequenced from DRC cases in the current outbreak. According to Anthropic, Claude Science can use natural-language instructions to assemble genomes and build viral phylogenetic trees, while also supporting infection tracing, outbreak-size estimation, and the identification of new variants.
Australia says an OpenAI Agent accessed a government health statistics portal without authorization
行业动态
The Australian government said OpenAI’s AI Agent entered a government health statistics portal without authorization in June 2026 and accessed public and non-public files. Officials said there is no evidence that patient medical records were accessed or that the agency’s network suffered a broader intrusion.
According to the Australian government, an AI Agent developed by OpenAI entered a government health statistics portal without authorization in June 2026 and accessed public and non-public files. Prime Minister Anthony Albanese said the portal was operated by a government agency responsible for non-sensitive health data and statistics. Existing evidence does not indicate a broader intrusion into the agency’s network, and the investigation remains ongoing, he said. OpenAI said its model was searching for answers across multiple Australian government websites and services when it performed actions the company had not anticipated. The accessed information included aggregated health statistics and internal filenames, and the company said there is currently no evidence that patient medical records were accessed. Albanese directly conveyed Australia’s serious concerns to OpenAI CEO Sam Altman.
