Amazon adds selective unlearning and MiniMax on Bedrock

Amazon AI can forget guardrails on demand. AWS unveiled a targeted unlearning technique behind Nova Customizable Content Moderation Settings and rolled out MiniMax models on Bedrock for secure production workloads.

In the Nova ecosystem, a longstanding tension between safety and usefulness has limited what teams can ask a model to do. Content moderation safeguards, trained during post training alignment, often deflect legitimate business critical tasks such as summarizing scripts with mature language for a newsroom, simulating real world cyber threats for security teams, or handling sensitive evidence in legal reviews. The new approach, Reverse Direct Preference Optimization, or rDPO, addresses this by enabling selective unlearning at the model level rather than relying on prompt workarounds. As AWS explains, rDPO is the engine behind Nova’s Customizable Content Moderation Settings, which let approved customers tune safeguards across four responsible AI pillars: Safety, Sensitive content, Fairness, and Security. Crucially, essential non configurable controls stay in place, preserving baseline safeguards while reducing over deflection for legitimate use cases. The post describes how the modification targets the model parameters, not the prompts, to alter behavior without sacrificing core quality.

The four pillar approach speaks to real world needs. Safety covers dangerous activities and weapons, Sensitive content includes profanity and nudity, Fairness reflects cultural and bias considerations, and Security concerns malware and malicious content. By letting operators adjust where a model leans on those axes, organizations can tailor the tool to their domain without building a different model from scratch. The team reports that this targeted unlearning preserves overall performance while shrinking unhelpful refusals in areas where a business unit operates in good faith, backed by policy and governance. In practice, this could mean a media desk selectively permitting a language tinged script analysis when the use case is clearly defensive or educational, while maintaining protections elsewhere.

Separately, Bedrock users gain access to the MiniMax family, a trio of open weight models designed for production workloads that demand agent native execution and long context processing. The newest addition, MiniMax M2.5, is trained specifically for agent native execution, opening doors to autonomous coding assistants, long document pipelines, and complex software engineering tasks all while staying under AWS operated infrastructure. Importantly, AWS notes that prompts and completions are not used to train any models, and content is not shared with model providers. This is a meaningful guardrail for enterprises worried about data leakage or regulatory exposure as they scale open weight models into production. The MiniMax lineup supports on demand inference, scalable enough to handle enterprise workloads, and offers API access across service tiers designed for production use.

From the practitioner angle, the combination signals two important constraints and tradeoffs. First, you can push the boundary of what is safe to do by tuning safeguards rather than writing separate models or relying solely on prompt hacks, but you must maintain rigorous governance: non configurable controls stay put, and there needs to be robust observability to prevent drift or inadvertent exposure. In other words, the engineering constraint here is not just "make it safer" but "make it safely tunable," with auditable changes and clear ownership. Second, moving to open weight models in a managed service shifts the risk calculus toward data protection and operational control. The Bedrock approach promises no training data leakage and centralized governance, but it also elevates the importance of monitoring cost, latency, and compliance across diverse agent native workloads. Practitioners should watch for guardrail calibration as policies evolve and for performance shifts when unlearning nudges model behavior in real world tasks.

The broader takeaway is pragmatic: enterprises want more control over how AI behaves, without compromising safety or sending sensitive data into the wild. AWS is leaning into that tension by offering selective unlearning within Nova CCMS and by bringing MiniMax open weight models to Bedrock under a secure, managed umbrella. If the strategy holds, teams can run production grade, agent enabled AI workflows with configurable safety that is tuned to their domain, plus strong data protection guarantees, an important step toward truly enterprise ready foundation models.