OpenAI announced its upcoming Astra model, Anthropic introduced Claude Fable 5.1 and Claude Mythos 5.1, while Google unveiled Gemini 3.8 Flash and Gemini 3.8 Flash Cyber.
Highlights
- OpenAI said its upcoming Astra model has become its first to reach the “Critical” cybersecurity capability tier under the company’s Preparedness Framework.
- On ExploitBench, Astra achieved a perfect score on known vulnerabilities and, according to OpenAI, discovered two genuine zero-day vulnerabilities in a separate internal evaluation.
- OpenAI said parts of Astra’s release were delayed by several weeks to develop and test stronger safety protections.
Leading global technology companies OpenAI, Anthropic and Google have announced new artificial intelligence models and updates, bringing new capabilities while also triggering fresh debate around AI safety and monitoring.
OpenAI announced that its upcoming AI model Astra has reached the “Critical” cybersecurity capability tier under its Preparedness Framework.
The framework classifies OpenAI models into capability tiers based on what they can do and determines the safety controls required before further development or deployment.
OpenAI said Astra, when provided with appropriate tools and access, can identify previously unknown security vulnerabilities and develop ways to exploit them across well-protected systems without requiring a person to guide every step.
During testing on ExploitBench, a public exploit-development benchmark, Astra achieved a perfect score on known vulnerabilities. OpenAI said the model also identified two genuine zero-day vulnerabilities in a separate internal evaluation designed to avoid training-set contamination.
The company said the vulnerabilities were being disclosed to the relevant software maintainers.
OpenAI also said it delayed parts of Astra’s release by several weeks while developing and testing stronger protections. Access to its most advanced cybersecurity capabilities will initially be restricted to vetted users, while broader defensive access will be available through a programme called Daybreak Blue.
OpenAI CEO Sam Altman described Astra as a significant step forward in capabilities and alignment and said the company has been slowing development when necessary to provide more time for safety work.
However, Astra has triggered debate following reports that the model uses a technique called “recurrent depth.” Concerns have been raised that the approach could shift some reasoning into internal mathematical computations, making it harder for safety researchers to monitor the model’s reasoning.
OpenAI Chief Scientist Jakub Pachocki pushed back against some of those concerns, saying reports about Astra’s architectural changes were confused and that the company continues to value chain-of-thought monitoring.
Meanwhile, Anthropic launched Claude Fable 5.1 and Claude Mythos 5.1, claiming top benchmark scores for coding and scientific research. It also said prices for typical workloads have been reduced by around 25%.
Google announced Gemini 3.8 Flash and Gemini 3.8 Flash Cyber, alongside agentic video capabilities and Google Pics, a new image creation and editing tool.
