Created automatically with AI: the texts and images in this edition are AI-generated, based on the linked original sources. Errors may occur.
Anthropic is expanding Claude with new work features and tightening its usage rules. At the same time, AI-generated mathematics, safety culture, and the selection of models for search services are coming into focus.
AI-generated image
01The Decoder
Claude creates dashboards and animated explainer videos in new betas
Anthropic has introduced two beta features for Claude: Dashboards are intended to turn data sources such as BigQuery and Snowflake into continuously updated overviews through text prompts. Motion creates animated explainer videos from text, diagrams, and images; according to the provider, the results can be edited as code and exported as MP4 files. Dashboards is available to paying users, while Motion is available to Team and Enterprise customers. Docs, Slides, and Design are now also open to free accounts.
Why it matters
The features expand Claude with tools for data visualization and media production. This takes the application beyond text and chat tasks and into practical workflows.
Anthropic clarifies rules for Claude—including for autonomous applications
Anthropic’s updated usage rules take effect on November 12, according to the company. Among other things, they prohibit deceptive campaigns using fake accounts, the use of Claude for weapons components, and certain surveillance practices. For systems that use Claude to perform physical actions and could cause injuries, Anthropic requires a qualified supervisor who can intervene. Also new is a ban on sustained, purposeless cruel behavior toward the models; ordinary frustration and testing are exempt.
Why it matters
The rules define which uses of its models Anthropic excludes and what supervision it requires for applications with possible consequences for people. They therefore also affect developers and operators of AI-powered systems.
OpenAI’s AI manuscripts spark dispute over scientific practice
According to the reports, OpenAI published 722 AI-generated manuscripts on mathematical results; according to Decoder, three papers were withdrawn the following day because of a sign error. The Association for Human Mathematics criticizes the bundled publication and calls on researchers to stop collaborating with OpenAI. At the center of the debate is whether results can count as progress if proofs have not been independently checked and worked out in a way that people can understand.
Why it matters
The dispute is about not only the correctness of individual AI results but also review, comprehensibility, and recognition in research. The debate could influence how AI-assisted mathematical work is published and evaluated.
Three dismissed OpenAI safety researchers dispute the company’s account
In an open letter, three dismissed safety researchers deny that they improperly disclosed confidential information and warn of a chilling effect on internal safety work. OpenAI told TechCrunch that an investigation had found a pattern of policy violations; the dismissals were not because of safety concerns they had raised. The researchers also deny any connection to a media report. OpenAI has not officially responded to the letter; an internal memo defended the decisions.
Why it matters
The case raises questions about how AI companies regulate the handling of confidential information and exchanges with external reviewers. The differing accounts have not yet been independently resolved.
Ecosia replaces Mistral with open-weight models, including Chinese systems
According to founder Christian Kroll, Ecosia is moving away from Mistral and plans to rely more heavily on open-weight models. Kroll told Politico that he made the switch because of technical problems, outages, and performance shortcomings. Through provider Melious, the search engine uses Qwen, GLM, and Kimi, among others. According to Kroll, this cut operating costs by about half and improved quality and performance. Experts warn that political biases in training data cannot simply be eliminated.
Why it matters
The switch shows that providers of AI services weigh costs, performance, infrastructure, and strategic independence against one another. The use of Chinese models also raises questions about censorship and influences from training data.
Anthropic launches cyber program for critical infrastructure and open source
Anthropic has announced a long-term cyber program intended to support operators of critical infrastructure and open-source projects. At launch, it includes a service that regularly checks registered open-source projects for vulnerabilities using Claude. According to the company, the reports are generated without human review and may contain errors. Selected service providers for critical infrastructure are also to receive access to models, technical experts, and threat research. Anthropic says it plans to expand the offering gradually.
Why it matters
AI-assisted security checks could give maintenance teams additional indications of vulnerabilities. Because Anthropic says the automated reports are shared without review and may be inaccurate, interpreting them remains an important task.
Cisco announces AI phone assistant and real-time translation for Webex
Cisco has unveiled new AI features for its telephony and collaboration platform at the WebexOne conference. A phone assistant is expected to answer calls in the first half of 2027, identify callers’ needs, and route calls, schedule callbacks, or answer questions using a knowledge base. The Webex Translator Agent is intended to translate phone calls in ten languages in real time. Cisco also announced cloud-based AI features for existing on-premises installations, including transcriptions and call summaries for Contact Center customers.
Why it matters
The announced features combine voice AI with business telephony and are also intended to include existing local systems. Some of the new features have only been announced for the future.
AI benchmarking platform Arena raises $200 million and expands evaluations
Arena, operator of the LMArena model leaderboard that grew out of a UC Berkeley research project, has closed a $200 million funding round, according to TechCrunch. The company was valued at $3.1 billion in the deal. Users compare models on the platform based on their responses; Arena also provides analyses for companies and AI labs. A new Alignment leaderboard evaluates, among other things, unauthorized actions, incorrect source attribution, and falsely claimed task completion.
Why it matters
The new category expands model comparisons to include behaviors that conventional performance benchmarks capture only to a limited extent. Such evaluations nevertheless remain dependent on the selection and design of the tests.
Automatically summarized with AI and checked against the source texts retrieved. Original sources are linked for every story; errors may still occur. An editorial overview from Medienfuchs.
