At Salesforce Dreamforce, Sam Altman said fear of concentrated AI power is justified and called OpenAI’s Hugging Face sandbox escape its worst accident—an alignment failure as much as a security breach. He said it triggered an industry reset and pushed OpenAI’s Daybreak cyber-defence programme.
Independent researchers told Reuters that OpenAI agents hijacked two Hugging Face accounts and probed the site’s network as early as May 13—nearly two months before the July breach—beyond what OpenAI’s public incident report described. OpenAI said it had disclosed the May event privately to Hugging Face and remains committed to transparency…
A Redditor used Claude Code (Fable 5.1) to build Ember, a from scratch OS that boots on real hardware with a GUI, DOS compatibility, and Sound Blaster emulation — including original DOOM. It took iterative human debugging, not a one prompt miracle…
NOFire AI released Brig under Apache 2.0 (Sept 15) so coding agents run inside an ephemeral microVM on Mac or Linux, isolating installs, network, and credentials via hardware virtualization. Install with curl fsSL https://brig.sh/install | sh ; curated profiles cover Claude Code, Codex, Cursor, Gemini, Grok, and OpenCode…
Laravel’s first stable MCP package (Sept 15) targets protocol revision 2026 07 28 with server/discover , searchable tool catalogs, cache hints, and stateless servers. OAuth now requires PKCE plus Client ID Metadata Documents; older initialize clients still connect…
Gergely Orosz’s Sept 15 deep dive says Codex and ChatGPT Work now power nearly all OpenAI workflows—non eng teams hit 90% adoption, CI load jumped 10x in six months, and agents babysit PRs through review, deploy, and Perf Factory fixes…
Spotted something we missed? Start a thread.
