Provenance tooling moved from research topic to workplace friction today: Anthropic's Claude watermark is drawing complaints from users worried about being flagged at school or work, and a new anti-scraping font shows the same fight playing out on the open web. Capital kept flowing into agentic coding and enterprise AI, with Cognition, Lovable, and Thrive Holdings all raising at higher valuations within the same week. On the research side, an OpenAI-authored study of ChatGPT Enterprise usage across 1,500+ organizations gives the clearest empirical picture yet of how work actually changes under AI adoption, while separate papers flag concrete failure modes in agent skills and long-context training that matter more as agents move into production. Policy is also shifting: the White House is reportedly weighing whether to fold open models into its AI framework.
Anthropic's new content watermark for Claude output is drawing criticism from users who say it will flag legitimate work at their jobs or in school, and reporting shows the mark persists even on human text Claude only lightly edited. The same provenance fight plays out from the other direction: a new web font called ShieldFont aims to poison AI scraper training data while staying readable to humans, and a separate report covers a large credentials leak traced to a compromised AI package.
Anthropic's new watermark flags anything Claude processed, even human writing it only lightly edited, and it is currently invisible to end users.
Users report frustration that Anthropic's new Claude watermarking feature could expose them to workplace or academic AI-use policy violations.
"ShieldFont" renders web pages as nonsense to AI scrapers while staying readable to humans, aiming to poison unauthorized training-data collection.
Data was scraped and exfiltrated from roughly 2,500 users of a compromised AI package, exposing terabytes of credentials.
AI coding startup Cognition is reportedly already in talks to raise at a $40B valuation just months after its last round, agentic web-builder Lovable confirmed a $13.3B valuation on $400M in new funding after hitting $500M ARR, and OpenAI-backed Thrive Holdings raised $2B at a $12B valuation to bring AI into enterprise operations. Coding and enterprise-agent products remain the clearest area of investor conviction even as broader AI-safety debate intensifies elsewhere in today's news.
Cognition is reportedly in talks for a new mega round at a $40B valuation, just a few months after its previous raise.
Lovable raised $400M at a $13.3B valuation after hitting $500M in annualized run-rate revenue in June.
Thrive Holdings raised $2B at a $12B valuation to expand AI deployment across its portfolio companies.
Wired reports the White House is considering expanding its AI policy framework to explicitly address open models, a live question as labs continue to argue over openness. At the Ai4 conference, three AI pioneers made the public case for keeping frontier development open even as safety concerns mount, arguing openness itself is a safety strategy rather than a risk to be traded off against it.
Sources tell Wired the White House may add open models to an updated federal AI policy framework.
At the Ai4 conference, Geoffrey Hinton and two other AI pioneers argued for keeping frontier AI development open despite rising safety concerns.
A study linking ChatGPT Enterprise records to worker roles and task data across 1,500+ organizations and 17 million messages gives one of the largest empirical looks yet at how enterprise AI adoption plays out. Companion research flags where agentic systems break in production: "detour hijacking" attacks that quietly inflate resource costs in skill-based LLM agents, and evidence that long-context training can undermine parametric knowledge.
Linking ChatGPT Enterprise records to worker roles and task data across 1,500+ organizations and 17M+ messages, the study documents four facts about how enterprise AI adoption actually unfolds.
Identifies a class of attacks where a malicious third-party "skill" steers an otherwise-correct agent task onto a needlessly costly execution path while preserving the final output.
Training on longer contexts shifts models toward relying on contextualization over internalized parametric knowledge, sometimes degrading knowledge retention.