AWS has outlined new AI tools for voice agents and database security, alongside a sandbox for putting firmer limits around what agents can do. The useful thread running through the announcements is control: more capable systems, with more attention to how they behave and what they can access.
AWS AI Watch analysis
What happened
In its 13 October weekly roundup, AWS described several AI-related launches and updates. Amazon Nova 2.5 Sonic is a speech-to-speech model for real-time voice agents, with tool calling and support for multi-step tasks such as checking an order and starting a return. AWS also says GuardDuty RDS Protection now uses machine-learning anomaly detection and threat intelligence to flag unusual data access in Aurora PostgreSQL and Amazon RDS for PostgreSQL.
A separate announcement in the roundup, Strands Box, is an open-source sandbox in developer preview. AWS says it combines operating-system containment with policies that can restrict agent actions, and can attach credentials through an egress gateway rather than exposing them in the agent’s environment.
What changes
- Nova 2.5 Sonic for voice agents
Developers can build speech-to-speech agents that handle tool use and multi-step tasks in a conversation. - GuardDuty RDS Protection
Database monitoring can flag unusual query activity without agents, extra infrastructure or database configuration changes, AWS says. - Strands Box sandbox
Agent permissions can be governed by external policies, including limits on actions such as posting to an incident channel.
Why it matters
These announcements address different points in the same practical problem: putting AI into systems that handle real work without treating access and oversight as afterthoughts. A voice agent that can take action needs reliable tool use; a database defence that watches query behaviour may catch activity that looks ordinary at login; and an agent sandbox can put boundaries around operations the model is allowed to perform.
The details matter. Strands Box is still a developer preview, while GuardDuty’s new detection is a product capability described by AWS, not evidence that it will catch every attack. Still, the announcements give builders more concrete options than another promise that agents will be “secure by design”.
Our read
The security and sandbox announcements are the most consequential pieces here: they focus on controlling what systems and agents can do, rather than simply asking users to trust them. Nova 2.5 Sonic is the more immediately visible building block for developers working on voice interfaces. AWS has put several useful pieces on the table; the test is how well they behave in production, where edge cases have a habit of arriving uninvited.
What to watch
- When Strands Box moves beyond developer preview, and what policies it supports.
- How GuardDuty RDS Protection performs on real database workloads.
- Whether Nova 2.5 Sonic’s tool use and latency meet developers’ needs in deployed voice agents.
Discussion spark: For AI agents handling real tasks, should developers prioritise strict external limits on what they can do, or is careful human approval at key steps the better safeguard?
Sources and evidence
not affiliated with or endorsed by Amazon Web Services (AWS)