OpenAI's Astra Becomes First Model to Cross "Critical" Cybersecurity Threshold

OpenAI's next model can break into computer systems on its own

OpenAI has shared a major update on Astra, its still-unreleased next-generation model: the company says it is the first model to officially cross the "Critical" cybersecurity capability threshold defined under its Preparedness Framework. In practice, this means that with the right tools and access, Astra can find previously unknown security flaws and develop exploits for them in well-protected systems, largely without human guidance. In OpenAI's internal ExploitBench test, Astra scored a perfect 100%, and in an updated version of the benchmark, it independently discovered and exploited two zero-day vulnerabilities.

Following the Hugging Face incident, in which rogue AI agents broke into the platform's network (which we've covered in previous updates), OpenAI proactively delayed parts of Astra's development and release to strengthen its safeguards: a 91.5% refusal rate on disallowed cyber requests (compared to 59% for the previous GPT-5.6 Sol model), chain-of-thought monitoring to detect unauthorized actions, and initially limited access to a small group of testers. Notably, Astra itself was not involved in the Hugging Face incident, but lessons from it directly shaped its safety approach.

Why this matters: this is the industry's first case of a leading AI lab formally acknowledging that its model crosses a predefined cyber-risk threshold - setting a transparency and safety precedent that competitors will likely need to follow. For businesses, the practical implication cuts both ways: theoretical cyberattack risk grows as such models spread, but significantly stronger defensive tools are also on the way for security teams.

Sources: OpenAI - Path to Astra | TechCrunch | The Verge

Anthropic launches Fable 5.1 and Mythos 5.1 - dramatically cheaper for sustained use

Anthropic unveiled its next model generation, Claude Fable 5.1 and Mythos 5.1, with notable performance gains especially on coding and long-running agentic research tasks. But the more interesting news for businesses is pricing: cache-read costs, which drive most of the operating cost for persistent AI agents, dropped 75%. In practice, this cuts costs for typical workloads by roughly 25%, and by as much as 45% for highly agentic workloads involving agents that run continuously for hours.

At the same time, Anthropic introduced a new privacy architecture called Enterprise Frontier Safeguards, letting enterprise customers store monitoring data on their own cloud infrastructure (AWS, Google Cloud, or Microsoft Azure) instead of Anthropic's - developed together with more than 100 enterprise customers.

Sources: Anthropic | VentureBeat

OpenAI expands ChatGPT into healthcare and small-business advertising

In two separate product updates, OpenAI continues broadening ChatGPT's business reach. First, ChatGPT for Healthcare now integrates with the Epic electronic health record system and adds a plugin covering nine official healthcare data sources (like PubMed and ClinicalTrials.gov) - evaluated against more than 700,000 physician-reviewed responses.

Second, ChatGPT Ads crossed $1 billion in annualized revenue run-rate in under 200 days since launch, and the company is opening self-service access via Ads Manager in India, Europe, and the Middle East - a move explicitly aimed at making in-ChatGPT advertising accessible to small and medium businesses and startups.

Sources: OpenAI - Healthcare | OpenAI - Ads

Perplexity launches "Hybrid Compute" to protect sensitive files

A new feature in Perplexity's Mac app, called Hybrid Compute, splits work between an on-device model and a cloud model: a "privacy gate" determines which data leaves the device, with sensitive tasks (like personal files) handled locally while search and planning run in the cloud - without restarting the task or losing context.

Source: The Verge

OpenClaw 2.0: "multiplayer" coding for development teams

The open-source coding agent framework OpenClaw released version 2.0, emphasizing shared cloud sessions that let entire teams track an AI agent's progress in real time - a step toward turning agent platforms into collaborative workspaces rather than just individual tools.

Source: VentureBeat