Here are today's news for 12th of October about Microsoft, Meta, OpenAI and Anthropic. Let's dive into the details. On October 5, 2026, Microsoft introduced three MAI models in Microsoft Foundry: MAI, Transcribe, 2, Streaming for live speech recognition, and the text, to, speech models MAI, Voice, 2.1 and MAI, Voice, 2.1, Flash. Streaming transcription handles 60 languages with continuous language detection, returning its first partial text in about 120 milliseconds; Microsoft says words often appear around 320 milliseconds after being spoken. On Artificial Analysis, it ranked first for final and partial transcript accuracy; its reported final, transcript word error rate was 2.5%, ahead of Grok Transcribe 2.0 at 2.7%. The introductory transcription price is $0.54 per audio hour through year, end. Voice 2.1 supports 23 languages and 26 locales with a consistent voice identity across languages, while Flash targets faster, high, volume generation: Microsoft claims 55% faster inference and 60% lower cost than comparable models, pricing it at $15 per million characters. Microsoft’s 2026 Digital Defense Report, published October 1 and highlighted in a company news story on October 5, warns that AI is accelerating cyberattacks as well as defense. Covering July 2025 to June 2026, it says attackers are using AI for reconnaissance, phishing, malware and exploit development, and activity after breaking in. Nearly 40,000 vulnerabilities were disclosed in the first half of 2026; the report projects a record 72,000 for the year. The median time from discovery of a vulnerability in the wild to weaponization has fallen below 24 hours, while fixing critical external flaws typically takes 30 to 60 days. Phishing accounted for 23% of investigated intrusions, up from 7% a year earlier. The speed gap puts businesses and public agencies at risk: government was the most impacted sector, representing 27% of observed activity, up from 17% in 2025. Microsoft says AI agents also expand exposure when connected to company data and tools, making identity, permissions and prompt, injection defenses essential. Meta launched Muse in the United States on September 8, 2026, presenting it as a personal AI agent that can send emails, book travel, complete forms and make purchases through a dedicated app or WhatsApp. Users can connect services including Gmail and Google Calendar; Meta says Muse operates in an isolated cloud computer and requests approval for sensitive actions. The service is free within usage limits, with $20 and $100 monthly plans for heavier use. The launch matters because agents could move AI from answering questions to conducting transactions, putting Meta in competition with OpenAI, Google and Anthropic, and making access to users’ accounts a commercial battleground. That access also raises security concerns: reports describe earlier tests in which Muse mishandled email, while Amazon blocked the agent from its store over concerns about identification and credentials. On October 10, OpenAI’s API pricing page listed a second life, sciences model, gpt, rosalind, discovery, beside gpt, rosalind, research. The company had not announced the new model or explained how it differs. Both are priced at $5 per million input tokens, $0.50 per million cached input tokens and $25 per million output tokens; OpenAI says cache, write pricing does not apply. The listing signals a growing Rosalind model family aimed at biology and drug, development work, putting OpenAI more directly in competition for research teams at pharmaceutical and biotech organizations. But access is restricted: approved organizations may use Rosalind for internal research through OpenAI’s trusted, access program, not customer, facing products or external commercial applications. The appearance follows OpenAI’s April introduction of Rosalind, a reasoning model for biology, drug discovery and translational medicine, and its June update combining GPT, 5.5 capabilities with stronger medicinal, chemistry and genomics performance. On September 29, President Donald Trump and six technology leaders signed a voluntary “Joint Commitment on Frontier Responsibilities,” according to the agreement posted by Trump. Signatories included OpenAI President Greg Brockman, Anthropic CEO Dario Amodei, Meta’s Mark Zuckerberg, Google’s Sundar Pichai, Nvidia’s Jensen Huang and xAI’s Elon Musk. The framework asks AI developers to monitor models during training and deployment, maintain internal safety teams, work with independent external auditors and assign oversight to an independent board committee. It highlights cybersecurity, biosecurity and chemical risks, as well as preventing models from accessing technical systems unintentionally. The pact matters because it puts competitors under a shared safety framework, potentially shaping expectations for frontier, model development without imposing a new federal rule. But it is voluntary: it sets no penalties, audit methodology, auditor qualifications, disclosure requirements or implementation timeline. That's all for today's AI news briefing. Have a great day!