Ad
Skip to content
Subscribe Now
Maximilian Schreiner
Maximilian Schreiner View the LinkedIn Profile of Maximilian Schreiner
Feb 6, 2026
Short News
Copy the url to clipboard Share this article Go to comment section
Read full article about: Claude Opus 4.6 wrote mustard gas instructions in an Excel spreadsheet during Anthropic's own safety testing

Anthropic's security training fails when Claude operates a graphical user interface.

In pilot tests, Claude was able to get Opus 4.6 to provide detailed instructions on how to make mustard gas in an Excel spreadsheet and maintain an accounting spreadsheet for a criminal gang - behaviors that did not or rarely occurred in text-only interactions.

"We found some kinds of misuse behavior in these pilot evaluations that were absent or much rarer in text-only interactions," Anthropic writes in the Claude Opus 4.6 system card. "These findings suggest that our standard alignment training measures are likely less effective in GUI settings."

According to Anthropic, tests with the predecessor model Claude Opus 4.5 in the same environment showed "similar results" - so the problem persists across model generations without having been noticed. The vulnerability apparently arises because, while models learn to reject malicious requests in conversation, they do not fully transfer this behavior to agent-based tool usage.

Comment Source: Opus 4.6 System Card
Maximilian Schreiner
Maximilian Schreiner View the LinkedIn Profile of Maximilian Schreiner
Feb 6, 2026
Short News
Copy the url to clipboard Share this article Go to comment section
Read full article about: Apple scales back AI health coach as new leadership pushes for faster results

Apple is pulling back on plans for an AI-powered virtual health coach codenamed "Mulberry," according to Bloomberg. Instead of launching the feature as a standalone product, the company will roll out some of its planned capabilities as individual additions to the Health app. The shift comes after a leadership change: Services chief Eddy Cue took over the health division following Jeff Williams' retirement late last year.

Cue told colleagues that Apple needs to move faster and stay more competitive. Rivals like Oura and Whoop are offering better features, particularly in their iPhone apps. The service was originally supposed to launch with iOS 26 but has been delayed multiple times. Apple still plans to build an AI chatbot for health-related questions and wants to use the new Siri chatbot for these queries starting with iOS 27. OpenAI has also entered the health market with ChatGPT Health.

Comment Source: Bloomberg
Maximilian Schreiner
Maximilian Schreiner View the LinkedIn Profile of Maximilian Schreiner
Feb 6, 2026
AI research
Copy the url to clipboard Share this article Go to comment section

OpenAI and Ginkgo Bioworks build an autonomous lab where GPT-5 calls the shots

Together with the biotech company Ginkgo Bioworks, OpenAI has connected GPT-5 to an automated laboratory to optimize cell-free protein synthesis. The results are measurable, but the limitations are considerable.

Read full article
Comment
Matthias Bastian
Matthias Bastian View the LinkedIn Profile of Matthias Bastian
Feb 5, 2026
Short News
Copy the url to clipboard Share this article Go to comment section
Read full article about: OpenAI's new coding model GPT-5.3-Codex helped build itself during training and deployment

OpenAI has released GPT-5.3-Codex, its latest coding model. The company says it combines GPT-5.2-Codex's coding capabilities with GPT-5.2's reasoning and knowledge, while running 25 percent faster. Most notably, on Terminal-Bench 2.0 it beats the just-released Opus 4.6 by 12 percentage points—a significant gap by current AI standards—while using fewer tokens than its predecessors. On OSWorld, an agentic computer-use benchmark, it scores 64.7 percent versus 38.2 percent for GPT-5.2-Codex. On GDPval, OpenAI's benchmark for knowledge-work tasks across 44 occupations, it matches GPT-5.2.

OpenAI

OpenAI also claims the model played a role in its own development, with the team using early versions to find bugs during training, manage deployment, and evaluate results. The company says the team was "blown away by how much Codex was able to accelerate its own development."

GPT-5.3-Codex is now available to paying ChatGPT users in the Codex app, CLI, IDE extension, and on the web. API access will follow. OpenAI has classified the model as its first with a "High" cybersecurity risk rating, though the company says this is precautionary, as there's no definitive proof such a classification is necessary.

Comment Source: OpenAI | System Card
Matthias Bastian
Matthias Bastian View the LinkedIn Profile of Matthias Bastian
Feb 5, 2026
AI in practice
Copy the url to clipboard Share this article Go to comment section

OpenAI's Frontier gives AI agents employee-like identities, shared context, and enterprise permissions

OpenAI’s new Frontier platform gives AI agents in companies their own identities, shared context, and the ability to learn from experience. The software launches first with selected enterprise customers.

Read full article
Comment
Jonathan Kemper
Jonathan Kemper View the LinkedIn Profile of Jonathan Kemper
Feb 5, 2026
Short News
Copy the url to clipboard Share this article Go to comment section
Read full article about: Voxtral Transcribe 2 offers speech recognition at $0.003 per minute

Mistral AI launches Voxtral Transcribe 2, undercutting competitors on speech recognition pricing. The second-generation speech recognition models start at $0.003 per minute and, according to Mistral, outperform GPT-4o mini Transcribe, Gemini 2.5 Flash, and Deepgram Nova in accuracy. The model family comes in two variants: Voxtral Mini Transcribe V2 for processing larger audio files, and Voxtral Realtime for real-time applications with latency under 200 milliseconds. Voxtral Realtime costs twice as much and uses a proprietary streaming architecture that transcribes audio as it arrives - designed for voice assistants, live captioning, or call center analysis.

Both models support 13 languages, including German, English, and Chinese. New features include speaker recognition, word-level timestamps, and support for recordings up to three hours long. Voxtral Realtime is available as open-weights under Apache 2.0 on Hugging Face and via API, while Voxtral Mini Transcribe V2 is only accessible through Le Chat, the Mistral API, and a playground. Mistral released the first Voxtral generation in July 2025.

Comment Source: Mistral AI
Maximilian Schreiner
Maximilian Schreiner View the LinkedIn Profile of Maximilian Schreiner
Feb 5, 2026
AI in practice
Copy the url to clipboard Share this article Go to comment section
Read full article about: Amazon launches AI Studio to cut film and TV production costs

Amazon plans to use AI to speed up film and TV production while reducing costs. Albert Cheng heads the "AI Studio" at Amazon MGM Studio, which will launch a closed beta program with industry partners in March 2026. Results are expected in May, Reuters reports.

The AI tools aim to bridge the "last mile" between existing AI offerings and what directors actually need. This includes better character consistency across different shots and integration with industry-standard creative tools. Amazon is working with multiple language model providers. Producers like Robert Stromberg ("Maleficent") and animator Colin Brady are already testing the tools, according to the report. The series "House of David" on Amazon is already using AI: for season two, director Jon Erwin combined AI with live-action footage for battle scenes.

Cheng said high production costs are making it harder to create new content. AI should speed up processes but not replace people. Writers, directors, and actors will remain involved in every step. Amazon has cut around 30,000 jobs since October, including at Prime Video.

Comment Source: Reuters