OpenAI pauses Astra model development over critical cybersecurity risks

OpenAI has identified potential "critical" cybersecurity capabilities in its upcoming artificial intelligence model, Astra, leading the company to halt specific internal development processes. The San Francisco-based startup disclosed the finding on Friday, noting that the model’s advanced reasoning could potentially be exploited for digital attacks. Internal development paused The decision to pause specific workflows follows rigorous safety evaluations of Astra, which is designed to be a multimodal AI agent capable of seeing, hearing, and reasoning in real-time. OpenAI’s internal safety protocols categorize "critical" risks as those that could significantly assist in the creation of cyberattacks, including automated vulnerability research or the generation of malicious code that evades current defenses. Company officials stated they cannot currently rule out the possibility that Astra’s capabilities exceed safe thresholds for public release. The pause allows the engineering team to implement additional guardrails and "red-teaming" exercises to ensure the model does not provide actionable intelligence to threat actors. Global AI safety scrutiny This development comes as global regulators, including those in the UAE and the United States, increase pressure on AI developers to demonstrate safety before deploying frontier models. The UAE’s AI Office has previously emphasized the need for robust governance frameworks to mitigate risks associated with generative technology in…