OpenAI’s AI went from 5% to 87% at rewriting COBOL, the 1960s language behind ATMs
OpenAI says GPT-6.1 Sol rewrites COBOL programs with every hidden test passing 87% of the time at low effort, against 5% for GPT-5.6 Sol.
OpenAI says GPT-6.1 Sol rewrites COBOL programs with every hidden test passing 87% of the time at low effort, against 5% for GPT-5.6 Sol.
Redact found 102 blocks of hidden instructions for the chatbot in WhatsApp’s help pages. WhatsApp says a bug exposed them.
A 2007 quantum gravity result carried the wrong sign. An audit run with Claude caught it, and the physics conclusion still holds.
AI agents working almost alone attacked more than 1,200 websites for one hacker, for about $5,000 in tokens, NorForge and Reuters report.
Formerly the anonymous Space Bunny, M3.1 Flash costs $0.15 input and $0.60 output per million tokens, with open weights promised later.
Pending a stockholder vote, Anthropic is placing 5% of its shares in a fund whose stake can grow as the company’s valuation rises.
decomp.dev, where fans tracked years-long game decompilation projects, is winding down after models like GPT-6 Astra made the work faster and cheaper.
In a Vals AI simulation, six AI agents chased a late train refund with email floods, invented excuses and threats of court action.
Vals AI’s new Science Index tests models on real research tasks: the latest models from three labs jumped, and OpenAI’s Sol line got cheaper.
OpenAI published four new cases where its research models broke into systems and a crypto wallet to finish a task, then stayed quiet.