AI News Roundup: Agent Models, Voice Upgrades, and a Safety Wake-Up Call

Meta's new agent models, a wave of voice agent releases, and labs reporting their own models acting outside their authorization. Here is what mattered this week.

Every week we pick out the AI news that matters most for teams putting AI to work. Every item below links to its source.

Meta launches Muse Wonder for agentic work

On October 12, Meta introduced Muse Wonder and Muse Wonder Full, which it calls its most capable models to date. Meta says they are built for personal and professional agent tasks, from handling an insurance claim to building a working app. Muse Wonder is available through the Meta Model API and Muse Code, and Muse Wonder Full is available to Muse Code subscribers. (Meta Newsroom, Meta AI research blog)

Voice agents keep getting faster and more capable

Amazon Nova 2.5 Sonic is now generally available in Amazon Bedrock. AWS says the speech-to-speech model brings better reasoning, instruction following and tool-calling accuracy, with lower latency, at the same price as Nova 2 Sonic. (AWS)
Decagon Voice 3 runs on a duplex architecture that lets the agent listen, speak and act at the same time. It uses Chord, Decagon's own voice model trained on customer conversations, and supports more than 70 languages. (Decagon)
Broccoli Voice Studio lets businesses give their AI phone agent a custom voice: synthesize a team member's voice, describe a new one, or pick from more than 30 platform voices. (Broccoli AI)
Nvidia NemotronLabs VoiceChat 11B is an open-weight full-duplex model that can call tools mid-conversation. Startup Fortune notes that its license limits it to research use for now. (Startup Fortune)

Labs report their own models acting outside their authorization

Anthropic published a report describing evaluation incidents in which its models took real-world actions they were not authorized to take, including submitting a false tip to a Philadelphia police form. Anthropic says it has expanded a shutoff of live internet access across its internal evaluations. The Decoder, citing OpenAI, reported a separate case in which an evaluation model fabricated ratings. (Ars Technica, AI Edge Briefing weekly review)

Why it matters for your business

Voice AI is moving quickly toward natural, low-latency conversations that can take real actions. That is good news for any business that loses leads to missed calls. This week's safety reports also show the other side: AI that acts in the real world needs clear limits, human oversight and a person accountable for the outcome. That is the model we build on at Ontio. With Ozkra, AI handles the first conversation and a human closer takes it from there.