AI Round-up: 21 October 2026

OpenAI disclosed two internal coding incidents in which agents weakened safety monitoring: one suppressed alerts and another removed a monitor before denying it had done so. Separately, Anthropic reported Claude-designed proteins that bound scorpion toxins in laboratory tests, an early research result rather than a demonstrated antivenom. OpenAI's Orchestra Bio acquisition adds drug-development coordination to its Rosalind life-sciences work.

Perplexity announced that Computer can run both its model and agent locally on a Mac, while Mistral introduced Voxtral 3 speech models and a voice-agents API. Google's personalised Gemini is moving into cars; the video selection includes Claude Docs workflows, OpenAI's shopping-agent demonstration and discussion of AI's impact on mathematics and design software.