AWS Bedrock Unleashed: 18 New AI Models, AgentCore Upgrades, and Enhanced Security

During the re:Invent 2025 conference, AWS unveiled the most extensive expansion of its generative AI platform, Amazon Bedrock, to date—introducing 18 new open-weight models in a single release. In parallel, AWS also announced major upgrades to Amazon GuardDuty and AWS Security Hub, enhancing cloud security through more sophisticated AI-driven threat detection and risk correlation.

The latest Bedrock update reinforces its role as a “model marketplace,” enabling enterprises to select and switch between the most suitable AI models without rewriting application code. The newly added models include:

• Mistral AI: debuting Mistral Large 3, the company’s most advanced open-weight model optimized for long-context tasks, multimodal reasoning, and instruction reliability. The update also adds the Ministral 3 family—compact general-purpose models available in 3B, 8B, and 14B parameter variants.
• Popular mainstream models: additions include Google Gemma 3, MiniMax M2, and NVIDIA Nemotron.
• OpenAI: integration of OpenAI’s open-weight language model, gpt-oss Safeguard.
• Amazon Nova 2: AWS’s own Nova 2 models—Lite, Pro, and Sonic—positioned as industry-leading in performance-to-cost ratio for reasoning, multimodal processing, and conversational AI. AWS also introduced Amazon Nova 2 Omni, a single model capable of multimodal inference and image generation, available in preview starting today.

To give organizations greater flexibility, AWS has launched Nova Forge, the industry’s first open-training customization framework, granting access to pretrained checkpoints that can be fine-tuned with proprietary datasets.

Meanwhile, Nova Act achieves a breakthrough 90% reliability in browser-based workflow automation and is already adopted by companies such as Reddit and Hertz.

For building production-grade AI agents, Amazon Bedrock AgentCore now provides multiple preview features:

• Policy – allows teams to define clear behavioral boundaries for agents using natural language.
• Evaluations – includes 13 built-in evaluators to continuously assess correctness and safety.
• Memory – introduces episodic memory, enabling agents to learn from past interactions and grow progressively more capable.

AWS also demonstrated how various partners are leveraging its infrastructure:

• Adobe is reimagining creative and marketing workflows using Adobe Firefly’s text-to-image and text-to-video models trained on EC2 P5 and P6 instances. Adobe is also experimenting with Bedrock AgentCore to accelerate autonomous agent capabilities within its products.
• Sony has built two major platforms on AWS: an internal AI platform serving 57,000 employees (powered by AgentCore), and the Sony Engagement Platform, which connects electronics, PlayStation consoles, and media content. The latter processes data from over 500 sources through the Sony Data Ocean on AWS.
• WRITER now integrates Bedrock’s model catalog directly into its enterprise AI platform and has launched a new agent supervision suite—a unified control center offering detailed monitoring and centralized auditing to resolve trust and governance concerns at large scale.

On the security front, Amazon GuardDuty Extended Threat Detection is now generally available with expanded support for Amazon EC2 and Amazon ECS workloads. Previously focused on IAM, Amazon S3, and EKS environments, GuardDuty now covers virtual machines and container workloads as well. Trained at AWS scale, its AI and ML models correlate scattered signals—such as suspicious process creation, persistence attempts, or cryptomining activity—into a single high-severity alert mapped to the MITRE ATT&CK framework, accelerating incident response.

Additionally, AWS Security Hub has reached General Availability for its enhanced features. The service now offers near real-time risk analytics, automatically correlating signals from GuardDuty, Amazon Inspector, Macie, and other tools. With richer visualization and historical insights, Security Hub helps enterprises identify attack paths, understand how threats, vulnerabilities, and misconfigurations compound into risk, and prioritize remediation—addressing one of the most persistent challenges in multi-tool security operations.

Related Posts:

• AWS Unleashes Enterprise AI: Bedrock AgentCore & $100M Boost for AI Agent Development
• Amazon Launches Nova Premier AI Model with 1 Million Token Context
• AWS re:Invent 2025: Agentic AI, Marengo 3.0 Video Model, and Visa Payment Partnership