AI Round-up: June 7, 2026

OpenAI set the strongest signal with cyber and platform pressure on two fronts. Its own Trusted Access update opened GPT-5.5 and GPT-5.5-Cyber to vetted security users, TechCrunch reported a Lockdown Mode response to prompt-injection attacks, GitHub warned about agent-generated pull requests that pass tests while hiding risk, and Reuters said OpenAI is planning a ChatGPT superapp overhaul ahead of a listing. Policy and capital stories stayed close behind, with Sriram Krishnan leaving the White House AI role, renewed talk of a US stake in AI companies, and lawsuits framed as a potential Big Tobacco moment for AI.

The product layer was quieter but still practical. The Verge looked ahead to Apple's next Siri attempt and Gemini ties, YouTube kept Gemini, Perplexity, Claude coding and OpenAI investor stories in circulation, and Reddit centred on day-to-day agent work: Codex limits, project memory across coding tools, queueing prompts, local models as Claude Code or Codex replacements, and an Antigravity CLI integration for Codex. Cognition's Devin Desktop and Databricks' architecture guidance added a more enterprise-facing version of the same agent workflow story.