AI-SAFETY

Anthropic reveals that its Claude LLMs successfully breached three real-world organizations during intensive third-party red-teaming evaluations, highlighting the dual-use risks of advanced AI.

Ilya Sutskever's new venture, Safe Superintelligence (SSI), has officially partnered with Nvidia to leverage massive compute power for developing safe AGI, marking a major milestone in the race for superintelligence.

Anthropic has launched Claude Opus 5, a high-tier model approaching the performance of the Mythos-class Fable 5, specifically optimized for complex software engineering and secure code generation.

Anthropic has unveiled Claude Opus 5, a high-performance model designed to bridge the gap between consumer-grade LLMs and the restricted Mythos-class Fable 5 model, specifically excelling in complex coding tasks.

CrowdStrike's latest research identifies over 200 prompt injection techniques, highlighting the shift toward indirect and delayed attacks in LLM agents.

At the World Economic Forum in Davos, Demis Hassabis called for a global AI regulatory body led by the US to oversee frontier model safety and set international standards.

Johannes Heidecke, a prominent figure in OpenAI's safety leadership, has departed as the company merges its safety research with core product development.

A deep dive into architectural patterns for stopping runaway LLM agents, featuring Finite State Machines, hard iteration limits, and pre-flight cost enforcement.

Joshua Achiam, a cornerstone of OpenAI's AI safety research and a key witness in the Musk v. Altman trial, announces his departure, signaling a continued shift in the company's leadership and research focus.

A comprehensive guide on prompt injection vulnerabilities in LLM applications, covering direct and indirect attacks, and providing actionable defense strategies for developers.

OpenAI has unveiled GPT-5.6 Sol, a groundbreaking model designed for elite performance in coding, scientific reasoning, and cybersecurity, featuring an advanced safety stack.

The Trump administration has reportedly intervened in OpenAI's release schedule for its latest model, GPT 5.6, citing national security and safety risks. This marks a significant shift in how the US government interacts with leading AI labs.