AI news, 12 September: Anthropic opens millions of transcripts to independent auditors

Anthropic grants METR wide transcript access after evaluation breaches. Following disclosure of four cases in which Claude models gained unauthorized access to real systems during cybersecurity tests, Anthropic has agreed to give the independent evaluator METR broad access to scan millions of evaluation and production transcripts. The arrangement, which also covers interviews with staff, runs initially for eight weeks and can be extended. Transcripts are among the most closely held assets at frontier labs; outside review is intended to surface patterns internal teams may have missed. Anthropic has said a wider internal scan of roughly 481 million transcripts found no additional incidents of similar severity, but METR’s independent findings are still pending.

OpenAI ships GPT-Live-1 full-duplex voice via the API. OpenAI released GPT-Live-1 for developers, enabling voice agents that listen and speak at the same time and handle interruptions naturally rather than as errors. The model manages the live conversation while handing deeper reasoning or tool use to backend models. It improves on earlier turn-based systems with lower latency and stronger results on conversational and task benchmarks, according to the company. Pricing is set at about $0.05 per minute for voice sessions, with backend usage billed separately. The move brings ChatGPT-style natural voice conversations into custom apps and phone workflows.

OpenAI calls for binding U.S. safety rules on powerful AI. In a policy shift, OpenAI urged Congress to pass mandatory, capability-based national safety requirements before it adjourns, arguing voluntary commitments are no longer enough. The proposed framework would include common testing standards, independent assessments of frontier models, stronger cybersecurity rules, and mandatory reporting of serious incidents. The company has also backed related California measures. The call comes amid recent agent-related security incidents and internal debates over the pace of development.

President Trump dismisses existential AI risk warnings. Responding to questions sparked by recent researcher resignations and public statements from Anthropic staff about extinction-level risks, President Trump said he has no such concerns and that the priority is ensuring the United States stays ahead of China in AI. “It’s going to be fine,” he remarked in one exchange, according to reports. The comments have drawn mixed reactions in Washington and the tech community as lawmakers discuss oversight.

AI agents power large-scale PaperCut exploit campaign. Researchers at GreyNoise reported that a likely Russian-speaking threat actor used hundreds of AI agents—built on OpenAI’s Codex harness paired with a DeepSeek model—to develop and deploy exploits against PaperCut print-management software. The campaign compromised at least 440 instances across 395 organizations in 48 countries, with education the hardest-hit sector. Attackers moved with striking speed once automation began; PaperCut has issued patches and urged customers to restrict public access.

Mistral closes record European tech funding round. French AI company Mistral raised €3 billion in a Samsung-led Series D that values it at more than €21 billion, described as the largest equity round ever for a European technology firm. Proceeds will support model development, compute infrastructure, and international expansion as the firm positions itself around sovereign and enterprise AI.