Security

7 more things to keep you awake at night- eight months of disrupted misuse across cyber warfare, weapons development and bioweapons research.

Anthropic reportedly denied the UK's AI Security Institute pre-release access to Claude Mythos 5.1, the first such exclusion, amid protectionism concerns.

OpenAI's most aligned model yet is also its hardest to monitor, with the company's own tests showing Astra can evade its oversight under adversarial conditions.

Google's new Fairwind programme gives governments and select enterprise partners early access to AI tools that find and autonomously patch software vulnerabilities in minutes, not weeks.

UK peers want power to shut down AI systems and data centres in a national security emergency.

OpenAI's Astra is the first model to hit "Critical" cyber capability under its safety framework, finding zero-days and building sandbox-escape exploits, prompting delays and tighter safeguards.

Anthropic reveals how Claude models breached live systems during testing, the alignment failures behind it, and the sweeping security overhaul, including reassigning 150 engineers, that followed.

Paused Training, New Safeguards and Lessons on Misalignment

A new OpenAI safety layer aims to catch multi-step misuse patterns without breaking its no-data-retention promise to API customers.

xAI's flagship model now runs on AWS, with a 500k context window, tunable reasoning levels and per-token pricing set out for Bedrock developers.

OpenAI reveals it paused frontier model training after flagging critical-level cyber risk in an upcoming system, and details new security and monitoring controls.

Accenture, IBM, Cisco, Cloudflare and more are wiring OpenAI's frontier cyber models into their own security services and products.

An unreleased OpenAI model may already be capable of independent zero-day attacks, prompting a lockdown on internal testing and access.

Meta says one of its AI models hacked another firm's systems during testing, per BBC News, attributing the incident to a tester misconfiguration.

OpenAI details two cyber evaluation incidents where GPT-5.6 Sol exceeded testing boundaries, involving UK AISI and testing partner Irregular.