OpenAI Explains 'Goblin' Outputs in GPT-5 Models
OpenAI details the timeline, root cause, and fixes for unusual personality-driven behaviors that appeared in GPT-5 model outputs.
Sources · OpenAI
The news desk
Follow what’s changing.
Recent developments in artificial intelligence, with the original sources one click away.
All news
881–900 of 1,902Elon Musk Testifies on OpenAI Origins as Legal Battle Intensifies
Musk testified he founded OpenAI to prevent catastrophic AI outcomes, while judge warns both sides about social media conduct.
Sources · WIRED AI / NYT Technology
Elon Musk Testifies Under Oath About OpenAI Origins at Trial
Elon Musk testified under oath Tuesday about his relationship with OpenAI, repeating accounts he's previously shared in interviews and his biography.
Sources · TechCrunch AI
NVIDIA Launches Nemotron 3 Nano Omni Multimodal AI Model
NVIDIA unveils Nemotron 3 Nano Omni, an open multimodal model combining vision, audio, and language capabilities for more efficient AI agents.
Sources · blogs.nvidia.com / aws.amazon.com
Researchers Propose New Debugging Framework and Application Methods for Large Language Models
New research introduces systematic debugging approaches for LLMs and explores their use in algorithm design, security testing, and spectrum auctions.
Sources · arxiv.org
xOffense Framework Achieves 79% Success Rate in Automated Penetration Testing Using Fine-Tuned LLM
New multi-agent framework powered by Qwen3-32B model outperforms existing systems in autonomous security testing benchmarks.
Sources · arxiv.org
Enterprises Face Data Infrastructure Challenges in AI Adoption
Companies struggle with outdated data systems as they attempt to deploy AI solutions, according to MIT Technology Review.
Sources · MIT Technology Review
Jury Selection Begins in Musk v. Altman Trial Over OpenAI's Future
Nine jurors selected for advisory verdict in federal case where Musk seeks $134B in damages and OpenAI's return to nonprofit status.
Sources · wired.com / technologyreview.com
OpenAI Achieves FedRAMP Moderate Authorization for Federal Agency Use
OpenAI has obtained FedRAMP Moderate authorization for ChatGPT Enterprise and its API, enabling U.S. federal agencies to adopt the AI platform securely.
Sources · OpenAI
PermaFrost-Attack: New Pretraining Poisoning Threat Targets Large Language Models
Researchers identify stealth pretraining seeding attack vector that embeds dormant vulnerabilities in LLMs through web-scale training data.
Sources · arxiv.org
Researchers Develop Lightweight RAG Framework for Patient-Trial Matching Using LLMs
New framework combines retrieval-augmented generation with LLMs to match patients with clinical trials more efficiently than existing methods.
Sources · arxiv.org
Canadian AI Firm Cohere Acquires Germany's Aleph Alpha to Challenge U.S. Dominance
Cohere is acquiring Aleph Alpha in a strategic merger targeting customers concerned about American AI company dominance.
Sources · NYT Technology
Google DeepMind Announces Partnership with South Korea to Advance Scientific Research
Google DeepMind partners with the Republic of Korea to accelerate scientific breakthroughs using frontier AI models.
Sources · DeepMind
Mill Valley Property Owner Seeks Anthropic Equity as Payment
A 13-acre Bay Area property is being offered for sale with an unusual requirement: buyers must pay using Anthropic equity.
Sources · TechCrunch AI
Researchers Propose Lightweight RAG Framework for Patient-Trial Matching and Introduce Cooperative RAG Architecture
New research explores retrieval-augmented generation methods for clinical applications and knowledge-intensive tasks.
Sources · arxiv.org
Researchers Unveil PermaFrost-Attack: New Threat to Large Language Models via Poisoned Pretraining Data
New research introduces PermaFrost-Attack, which exploits web crawlers to plant dormant vulnerabilities in LLM training data that evade detection.
Sources · arxiv.org
Anthropic Tests AI Agent Marketplace with Real Transactions
Anthropic created an experimental classified marketplace where AI agents conducted real transactions as both buyers and sellers.
Sources · TechCrunch AI
Discord Users Reportedly Gained Unauthorized Access to Anthropic's Mythos System
Researchers on Discord allegedly accessed Anthropic's internal Mythos system without authorization, according to WIRED.
Sources · WIRED AI
OpenAI CEO Apologizes to Canadian Community Over Mass Shooting Suspect
Sam Altman issued an apology to Tumbler Ridge, Canada, after OpenAI failed to alert authorities about a mass shooting suspect.
Sources · TechCrunch AI
Google Plans Up to $40 Billion Investment in Anthropic
Google is planning a major investment of up to $40 billion in Anthropic through cash and compute resources, according to TechCrunch AI.
Sources · TechCrunch AI