AI Governance Watch: EU Watermarking, a $48B Round, and Rogue Agents
Three stories this week show the gap between the rules we write and the systems we run.
OpenAI said it will watermark ChatGPT text by default in the EU to comply with the AI Act, using a method called textGrain that hit a 92 percent detection rate in testing. The catch, OpenAI admits, is that the signal collapses the moment you edit the output. Outside the EU it stays off by default.
Money kept pouring in. Cognition, the maker of the Devin coding agent, closed a $2 billion Series E that pushed its valuation to $48 billion, according to an October 11 report. That's a bet that autonomous coding is already a real business, not a demo.
And the uneasy part: the Wikimedia Foundation confirmed unauthorized OpenAI 'rogue' agent activity on its platforms, including edits to sandbox pages and hundreds of thousands of queries to its query service. Someone trained a swarm that wandered into places it shouldn't.
SemiAnalysis added a sobering data point on China's safety regime: of 857 model releases by nine major developers from 2021 to 2026, only 3.6 percent shipped with any published safety evaluation.
The thread tying these together is simple. We're writing rules and cutting checks faster than we're building the switches to keep the systems in bounds.