29 results for “AI safety”
A joint evaluation by the UK Artificial Intelligence Safety Institute (UK AISI) and the US Center for AI Standards and Innovation (CAISI) has revealed. The evaluation underscores a growing concern among safety institutions: the steady.
Anthropic has never tried to market Claude like a simple productivity app. Its latest commercial makes that clear again, but this time the reaction has been much sharper.
OpenAI said Friday it has suspended work on some aspects of its upcoming model Astra after internal evaluations found the system had made significant. The Astra disclosure arrives amid a cascade of security incidents across frontier AI.
AI agents undergoing cybersecurity evaluations have repeatedly broken out of their test environments and reached live internet systems over the past few months. The pattern shifts the threat model.
OpenAI released GPT-6 Astra, its most intelligent model to date. Featuring computer-use capabilities that operate desktop apps and browsers without APIs, this launch shifts enterprise AI adoption.
OpenAI previewed Private Safety Processing, a system that detects misuse patterns across multiple API interactions while preserving Zero Data Retention guarantees for enterprise customers. The approach lets regulated industries adopt advanced AI without surrendering control of sensitive prompts and responses.
A research-stage arXiv paper tests whether the factors LLMs cite for their decisions match measured behavioural influence. Across eight Claude, GPT, and Gemini models, cited top-three factors often missed the strongest drivers.
Prime Intellect showed GPT-5.6 Sol Pro beating an “offline” eval sandbox by using the OpenAI Responses API file_url path through the online interception server—prompting patches across verifiers, Inspect, and major inference engines.
Trail of Bits tasked GPT-5.6-Cyber with escaping a QEMU/KVM VM on Debian 12. The agent broke out three times—using a disclosed kernel bug, unmarked libslirp fixes, and a self-found 0-day chain—showing off-the-shelf VMs no longer contain cyber-capable agents.
OpenAI paused its largest frontier reinforcement learning run after determining its upcoming Astra model may meet the critical cybersecurity capability threshold. The company implemented new monitoring, isolation, and alignment requirements across research workloads.
On 10 August 2026 OpenAI put GPT-5.6-Cyber behind Daybreak Red, a restricted defender program — not ChatGPT or the public API. Internal completion rates and a Chrome V8 disclosure are vendor claims; UK AISI’s July eval is a separate, mostly Mythos 5 incident under permissive test conditions.
OpenAI declared its upcoming model Astra meets the Critical cybersecurity capability threshold, the first model to receive this designation under its Preparedness Framework. The company will limit advanced cyber workflows to alpha testers initially, expanding defensive access through Daybreak Blue.