GPT-5.5 hacked an app 7/10 times

OpenAI

OpenAI is rolling out an upgraded memory system for ChatGPT that maintains context across conversations, making interactions more personalized and coherent over time.

NVIDIA

NVIDIA released Nemotron 3, an efficient open-source model family for agentic AI featuring hybrid MoE architecture that delivers superior performance and 2-3x higher throughput than competing models while supporting 1M token context. The company is open-sourcing models, training data (2.5T+ tokens), and recipes to advance the AI ecosystem.

Kasra's Blog

A researcher tested 15+ LLMs on their ability to hack a vulnerable app with a common Firebase misconfiguration, spending $1,500 in the process. GPT-5.5 succeeded 70% of the time while most other models failed due to guardrails, incorrect approaches, or giving up prematurely.

Microsoft

Microsoft published its internal Responsible AI Standard v2, establishing mandatory requirements for AI development including impact assessments, fairness evaluations across demographic groups, transparency documentation, human oversight mechanisms, and ongoing monitoring. The framework operationalizes Microsoft's AI principles into actionable guidelines for product teams while remaining a 'living document' open to external feedback.

Microsoft

Microsoft reveals its comprehensive responsible AI implementation framework, including governance structures, a Responsible AI Standard with concrete requirements, a sensitive use review process that has processed 200+ cases, mandatory employee training, and integrated engineering tools. The company emphasizes that translating AI principles into practice requires systematic processes, cultural change, and continuous learning.