AI Round-up: July 16, 2026

OpenAI supplied the clearest safety signal in this snapshot. Its GPT-Red thread framed automated red teaming as a way to harden GPT-5.6 against prompt-injection attacks, while its policy post tied AI safety to state and federal action. Anthropic pushed the same control theme from a different angle with new agentic-misalignment simulations, and Microsoft added a concrete supply-chain security case through its AsyncAPI npm compromise analysis.

The rest of the window centred on agent tooling and infrastructure. OpenAI and OpenAIDevs promoted Codex Micro, Perplexity described SPACE as the sandbox platform behind Computer, GitHub brought Copilot progress into Jira, Databricks surfaced Thinking Machines Lab's Inkling model and data-native agents, and DeepMind argued that scientific agents are creating a new validation bottleneck.