AWS AI Updates: September 19, 2026

1. AgentCore Runtime V2 Cut P75 Cold Starts From 30 Seconds to 2

AWS shipped a second-generation runtime for Amazon Bedrock AgentCore that changes two things about the serverless microVM model. Cold starts are now snapshot-based: the runtime prepares the agent environment once, snapshots it, and restores every new instance from that snapshot rather than replaying the full startup sequence, which holds P75 cold start at 1.9 to 2.0 seconds across container images from 200 MB to 2 GB. V1 took 5.4 to 30 seconds for the same image sizes, so startup time no longer scales with how much you pack into the image. Memory is also allocated on demand and reclaimed when idle instead of being held for the life of the session. Opt in by setting platformVersion to V2; available in us-east-1, us-east-2, us-west-2, eu-west-1, and ap-northeast-1. Source

2. Kimi K3 Is the First Open-Weight Model on Bedrock With Explicit Prompt Caching

AWS made Moonshot AI’s Kimi K3 generally available on Amazon Bedrock through serverless on-demand inference, reachable from every Bedrock region via cross-region inference. K3 is a 2.8 trillion parameter model with native vision and a 1-million-token context window, and Moonshot reports roughly 2.5x better scaling efficiency than K2. The notable Bedrock detail is that it is the first open-weight option on the platform to support explicit prompt caching, which matters for agent loops that resend a large system prompt on every turn. Source

3. AWS Continuum Now Validates Pentest Credentials Before the Test Runs

AWS added credential testing and accessible domain suggestions to Continuum, its penetration testing agent. Credential testing authenticates into the target application the way a real user would during test configuration, so an authentication failure surfaces before the run instead of after a full test cycle has burned. The same login pass captures every domain reached along the way and surfaces those as in-scope URL suggestions, which removes the manual step of enumerating what the application actually talks to. Available in all regions where Continuum for penetration testing already runs. Source