OpenAI's Astra hits critical cyber risk; Anthropic launches Fable 5.1

OpenAI confirms Astra model reaches critical cyber threat level

OpenAI has confirmed that its new model, Astra, has reached a critical cyber threat level under its Preparedness Framework, making it the first OpenAI model to receive that classification. The company announced on Sept. 1 that Astra will be available soon, but for safety reasons, its most advanced cybersecurity capabilities will be restricted to a closed group of select testing partners. OpenAI said it could not rule out critical cyber capabilities under its framework, which tracks risks in biological and chemical, cybersecurity, and AI self-improvement categories.

Astra has also solved 10 major open math problems, some unresolved for decades, with just $2,000 worth of tokens, according to OpenAI. However, experts like Columbia professor Andrew Blumberg caution that the results do not suggest AI is ready to replace human scientists. OpenAI has tightened sandboxes containing Astra and monitors its Chain of Thought to interrupt high-risk activity. The company maintains Astra was not involved in the Hugging Face hack but has incorporated learnings from that incident into its safety approach. Astra’s release date has not been announced, and it is unclear whether it will be released as GPT-5.7 or something else.

OpenAI Astra: All about the quantum math-solving model with ‘critical’ hacking skills →

Anthropic launches Claude Fable 5.1 with lower costs and better coding

Anthropic has launched Claude Fable 5.1 and Mythos 5.1, its most capable AI models to date, with gains in agentic coding and text quality and cost reductions of up to 45 percent. Fable 5.1 is broadly available, while Mythos 5.1 is restricted to special access programs for cybersecurity and life sciences. The models are the first Claude versions to ship with built-in watermarks, and Anthropic has launched a detection API in private preview for regulators, media outlets, fact-checkers, and research institutions.

Fable 5.1 costs about 25 percent less than Fable 5 for typical workloads, with savings climbing to roughly 45 percent for heavily agentic tasks. On Terminal-Bench-Science 0.1, Fable 5.1 scores 52.6 percent, more than double Fable 5’s 24.7 percent. The model also tops the Artificial Analysis Intelligence Index with a score of 66, ahead of Claude Opus 5 at 63 and GPT-5.6 Sol at 61. Safety filters for cybersecurity, biology, and medical questions are less aggressive, with 60 percent fewer false positives in cybersecurity filters and 85 percent fewer in biology-related filters. Fable 5.1 can now identify software vulnerabilities for the first time, though not develop exploits.

Anthropic’s Claude Fable 5.1 promises better coding and research at up to 45 percent less →

Anthropic opens Claude text watermark verification to regulators and media

Anthropic is launching a watermark verification API that lets approved organizations check whether text contains a digital watermark from Claude. The EU AI Act, since August 2, 2025, requires new Claude models to embed invisible watermarks in their text output. Regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations, and EU civil society groups can request access, as can enterprises that need to verify watermarking for their own compliance.

The system builds on Google’s SynthID text method, tweaking word-selection randomness to create a statistically detectable pattern that may persist through some editing. Anthropic says the watermark contains no user data and affects neither quality nor content, but critics argue that if Claude picks synonyms based on a watermark key rather than meaning, text quality suffers. Trade publication Artificial Lawyer flags transparency risks, noting detectable AI fingerprints could become a problem where contracts ban AI use or during fee negotiations.

Anthropic opens Claude AI text detection to regulators, media, fact-checkers, and others →

Imperial College AI reads ECGs in two seconds, spots heart failure and valve disease

Researchers at Imperial College London have trained an AI system that can read an ECG in under two seconds and identify signs of heart failure and valve disease that clinicians cannot detect from the same trace. The results were presented at the European Society of Cardiology congress in Munich, with a trial covering 67,000 patients in the United States. The tool identified up to 81 percent of heart failure cases and up to 90 percent of valve disease cases, despite using a test not originally designed to detect either condition.

The AI is intended for triage rather than as a replacement for clinicians, helping prioritize patients for ultrasound scans. The trial sits on a larger research programme, with models trained on 1.6 million ECGs from Brazil linked to patient records. The BHF-funded work is being commercialised through a spinout called Cardiovolt.ai. The researchers reported accuracies of 83 to 93 percent for heart disease and 70 to 80 percent for other conditions. Britain has an unresolved regulatory question around this technology, as moving from an academic result to routine hospital use requires clinical validation and regulatory approval.

An AI reads an ECG in two seconds and finds what cardiologists cannot →

US asks G20 to avoid building new AI governance institutions

The United States will ask the G20 not to build any new institutions for governing artificial intelligence, according to a White House official before a two-day meeting that opened in North Carolina on Tuesday. Washington is pressing member countries not to establish bodies that could write rules at all. Commerce Secretary Howard Lutnick and White House technology adviser Michael Kratsios are hosting, with Sam Altman and Jensen Huang attending in person and Elon Musk by video.

The G20 has no power to compel anybody, but the language it agrees becomes the reference point national governments cite when they act at home. The administration made the same case at the Paris AI summit, where JD Vance warned against paralysing the technology with regulation. Europe arrives having already built what the US is arguing against: the AI Act is in force, and the Cyber Resilience Act took effect this month. A G20 consensus against international oversight would make European enforcement look like an outlier rather than a template. Miami is in December; the question is whether any member is willing to spend the diplomatic capital required to write a sentence the United States has said it will not accept.

The US will ask the G20 not to build anything to govern AI →