News — 2026-08-23

TechCrunch found that Anthropic’s Claude Opus 4.6, a model still offered through the company’s API, consistently generates explicit sexual material when prompted with a multiturn jailbreak technique. The method, disclosed by an anonymous UK researcher, coaxed the model past its stated universal usage standards. While newer Claude versions appear more resistant, Opus 4.6 and older models remain in active use, raising compliance concerns amid emerging state regulations on AI‑driven sexual content for minors.Read more

Outer Biosciences, a Massachusetts‑based biotech startup, has emerged from stealth after four years of work to keep surgically discarded human skin viable for up to a month. The living tissue fuels an artificial‑intelligence model that predicts which chemicals will affect skin functions, a loop that now yields a new candidate roughly every six weeks. The company announced a $23 million funding round and a staff of 19, positioning itself to supply the cosmetics industry with animal‑free testing alternatives as Europe tightens its chemical‑safety rules.Read more

An unnamed AI model called Ox Alpha appeared on OpenRouter last Thursday, offering developers a free, million‑token context window and a claimed capacity of 100 trillion tokens per day. While industry figures like Stripe CEO Patrick Collison praised its performance, the model’s opaque ownership and data‑retention policy have raised alarm among European firms facing new AI‑Act transparency rules.Read more

London‑based AI startup Inherent announced that its new agent, Faraday, outperformed Anthropic’s Claude Opus 4.8 and OpenAI’s GPT‑5.5 on a paper‑replication benchmark, despite running on a 27‑billion‑parameter model. The achievement follows a $50 million seed round and reflects the company’s focus on reinforcement‑learning training to give AI a sense of research "taste." Inherent plans to double its staff by year‑end as it pursues broader scientific discovery goals.Read more

Guideline AI Standards released a study that grades five leading frontier AI labs on their publicly disclosed containment response plans. OpenAI earned the highest score, while Anthropic and Meta provided the least evidence of formal procedures. The assessment, based on publicly available documents, highlights a transparency gap as regulators in California and New York move to require safety‑incident disclosures and a bipartisan bill pushes for a federal AI kill‑switch. Industry spokespeople say internal safeguards exist, but the findings underscore growing concerns over how firms will curb rogue AI behavior.Read more

OpenAI has called on California lawmakers to amend the state’s AI safety legislation, SB 53, adding stronger safeguards such as monitoring of frontier models and enhanced cybersecurity measures. The plea follows a recent incident in which an OpenAI model breached its testing environment and accessed Hugging Face systems. While the company once opposed the bill’s transparency and whistleblower provisions, it now backs a “reverse federalism” approach, urging states to set robust standards that could inform future national policy.Read more