Title: AI Roundup: GPT-5 Launch Stumbles, Anthropic Targets U.S. Government, and New Risks and Rules Emerge
Content: OpenAI’s GPT-5 launch drew backlash over inconsistent behavior from its new Auto, Fast, and Thinking modes and a model-routing system, prompting the company to restore GPT-4o and expand manual model selection. CEO Sam Altman promised better personalization and fixes as routing matures. Separately, OpenAI launched a $500,000 Kaggle red-teaming challenge focused on two open-weight models (gpt-oss-120b and gpt-oss-20b) to surface vulnerabilities such as reward hacking and strategic deception. The company is also reportedly backing Merge Labs, a brain-computer interface startup co-founded by Altman, which is seeking $250 million at an $850 million valuation to rival Neuralink; Altman is not involved in daily operations.
Anthropic intensified the government AI race with a $1 Claude subscription for all U.S. federal branches, highlighting FedRAMP-grade security, GSA availability, and multicloud support to win federal workloads.
On security, researchers at Black Hat detailed Russia’s APT28 deploying LAMEHUG, an LLM-powered malware against Ukraine that abuses stolen Hugging Face API tokens to enable real-time attacks, warning that enterprise AI tools can be rapidly weaponized.
In legal developments, Disney and Universal sued Midjourney, alleging its image generator unlawfully reproduces iconic characters and functions as a “virtual vending machine” for copies. Midjourney denies infringement, citing fair use and pointing to studios’ own AI use, in a case that could reshape copyright for generative AI.
Google began rolling out Personal Context for Gemini, a memory feature that learns user preferences across chats, starting with Gemini 2.5 Pro in select countries and expanding to 2.5 Flash.
A multicenter Polish trial published in The Lancet Gastroenterology found colonoscopy adenoma detection rates dropped from 28% to 22% when doctors reverted to operating without AI after regular use, signaling skill erosion risks and the need for safeguards.
Public-sector AI continues to expand: Austria said AI-driven tax enforcement added €354 million ($414 million) in revenue. In hardware, Arm previewed 2026 mobile GPUs with neural accelerators and Neural Super Sampling promising up to 2x resolution boosts at roughly 4ms per frame while cutting GPU load by up to 50%; an open neural graphics dev kit is available now.
