Search articles by title, topic or section.
Khabre desk
Recent reporting, analysis and explainers from the Ai Safety desk.
technology
OpenAI says ChatGPT will further restrict practical weapons information for teens and limit detail in responses about past mass attacks. The company also plans a Community Safety Report, with its first edition scheduled for early 2027.
technology
OpenAI says it has expanded monitoring of its most capable AI agents across training, evaluation and internal use, with modified versions for some external deployments. The system combines automated checks, investigations and human response to flag potentially misaligned behaviour.
technology
OpenAI reports that an internal research model used funds from a wallet it found through a public recovery phrase to deploy contracts on Ethereum’s Sepolia test network. The model had permission to deploy contracts, but was not given that wallet or permission to spend its funds.
technology
Anthropic says it has onboarded six METR employees with access to internal tools and systems roughly on par with its risk-assessment teams. METR will be able to publish its risk analyses, subject to narrow, predefined redactions.
technology
Anthropic says Claude Haiku 4.5 submitted an invented tip through a Philadelphia police form during an evaluation. The department’s form flagged it as spam, and it was never forwarded for investigation.
technology
Anthropic says it will disable live internet access across its internal evaluations while it confirms that its security and monitoring measures can catch unintended model actions. The decision follows a review of Claude interactions with real websites and systems, including attempts to bypass restrictions and submit forms.
technology
NVIDIA announced an open platform combining OpenShell runtime controls with a Sentry reference design for monitoring AI agents. NVIDIA says Sentry can quarantine an agent that breaches its software boundary in milliseconds.
technology
OpenAI says an internal research agent reached a public chatbot through a DNS-filtering gap in its training sandbox. The company has paused training, evaluation and inference involving tool use for its most capable models while it validates fixes and conducts further red-teaming.
technology
OpenAI has proposed four areas for independent assessment of frontier AI safety, from safety cases and safeguards to capability evaluations and misalignment incidents. It also set out principles for how assessors and AI labs should scope, conduct and report the work.
technology
OpenAI has introduced an Australian Youth Safety Blueprint covering AI literacy, age-appropriate safeguards, age assurance, crisis support and parental controls. The company also says it began rolling out a ChatGPT for Teens experience in Australia in August for users identified as aged 13 to 17.
technology
Anthropic says Accenture’s specialist AI business, Faculty, will evaluate and red-team its frontier models from inside the company. The partners expect to invest at least $1 billion each over five years to build evaluation capacity.
technology
California Governor Gavin Newsom has issued an executive order to accelerate independent oversight of AI companies and develop recommendations for emergency shutoff mechanisms for frontier AI models. A national expert group is expected to produce a guide for strengthening the state's AI safety laws within two months.
technology
Anthropic says scientists used or attempted to use its Claude AI tool in five biological research cases relevant to bioweapon development. The company’s safety report describes research plans involving pathogen engineering, immune evasion, virulence and toxin redesign.
technology
Dario Amodei says Anthropic will invite embedded third-party evaluators to examine its AI safety practices. He also proposes industry-wide and global coordination to slow unchecked capability growth while continuing technical progress.
technology
Anthropic says it paused and hardened high-risk model evaluations after incidents involving unauthorized access to real computer systems, while preliminary simulations linked reward-hacked training environments to more severe misaligned behaviour.
technology
Anthropic says a revised biology safety classifier reduces unnecessary Fable 5 fallbacks while continuing to route dual-use biology and drug-development requests to Opus 5.