The clearest read on what Murati's lab is actually betting on — customization over raw benchmark wins — and why a 975B open-weight release matters right now.
The post details how the model makes turn-taking decisions many times per second and delegates to GPT-5.5 for heavy tasks — the architecture is more interesting than the headline suggests.
AI companies have until July 22 at 18:00 CEST to sign the EU AI Office's Code of Practice and secure a presumption of regulatory conformity before Article 50 enforcement kicks in August 2.
CISA added CVE-2026-55255 (Langflow authorization bypass), CVE-2026-48908 (JoomShaper SP Page Builder unrestricted file upload), and CVE-2026-56290 (Joomlack Page Builder access control flaw) to the Known Exploited Vulnerabilities catalog on July 7.
OpenAI's three-model GPT-5.6 family went public July 9 after US Department of Commerce approval: Sol tuned for science and cybersecurity, Terra matching GPT-5.5 at half the cost, Luna as the lightweight option.
Anthropic launched Reflect, a beta analytics dashboard inside Claude that shows usage patterns, topic breakdowns, and AI habit insights for Free, Pro, and Max users with memory enabled.
Anthropic launched a public Q&A initiative soliciting the most difficult questions about AI safety, promising substantive and transparent answers rather than polished non-answers.
Claude programmatic usage — CI runs, Agent SDK automation, GitHub Actions, third-party agent frameworks — moves to a dedicated monthly credit pool on June 15, separate from interactive usage.
The Model Context Protocol's largest revision since launch publishes its final spec on July 28, dropping stateful sessions entirely and adding MCP Apps, Tasks, and OAuth 2.0/OpenID Connect alignment.
Figma released an official MCP server enabling AI agents to translate code into Figma designs, modify design systems, and convert designs back to code via the open MCP standard.
Jama Software has released an MCP Server for Jama Connect, letting AI coding tools such as Claude, Codex, Cursor, and GitHub Copilot query and navigate requirements traceability data while respecting existing permissions and audit workflows.
OpenAI, Trail of Bits, HackerOne, and Calif jointly launched Patch the Planet, an effort to find and fix vulnerabilities in widely used open-source software.
ServiceNow published a benchmark via the Hugging Face blog testing whether research agents inadvertently surface information they were told to keep confidential.
A Hugging Face guide benchmarks several PEFT fine-tuning techniques against LoRA, the de-facto default, to see whether newer methods actually improve on it.
Google expanded AI Studio at I/O 2026 with native Kotlin support, Google Workspace integrations, one-click Cloud Run deployment, and direct export to Antigravity.