neuronlens.com

What It Is

NeuronLens is a platform for interpretability inside large models and runtime governance. Its core proposition is not merely to inspect prompts, outputs, or pipeline logs, but to read a model’s internal activations, sparse features, and concept signals, and turn them into deployable “Lenses” for safety, Agent supervision, domain intelligence, and model repair.

Core Capabilities

The platform is organized into four layers: Runtime Lens uses internal signals for real-time allow/block/reroute/review decisions; Concept Studio is used to search, inspect, and test concepts inside models; Domain Lenses target high-risk fields such as trading/risk, credit, cybersecurity, science, and healthcare; and Model Design Studio helps identify internal features behind recurring failures and apply targeted fixes. Its research roadmap includes sparse autoencoders, automated feature labeling, activation search, token-level heatmaps, feature families, contrastive fingerprints, steering, and reinspection. Agent Lens explicitly states that it can be integrated in two lines of code and supports LangChain, CrewAI, Claude Agents, AutoGen, OpenAI Agents, and more.

Pricing and Trial

The website only offers Request Demo and early access, and says it is working with a small number of research and enterprise teams. It does not disclose any free tier, public pricing, billing model, or payment methods. As a result, it looks more like an enterprise-customized or design-partner-stage product, and is not a good fit for individual users expecting instant self-service signup and a low-cost trial.

Pros and Cons

Its strengths are clear positioning and coverage of real-world risks such as prompt injection, RAG contamination, sensitive data leakage, tool-call hijacking, goal drift, and privilege escalation. It offers a closed loop from discovery to runtime control and then repair. Its research content is also relatively transparent, acknowledging the limitations of automated labels, sparse features, and steering. The main drawbacks are the lack of production case studies, a supported-model list, deployment performance data, privacy/compliance details, and security certification information. Chinese-language support is also not disclosed.

Who It’s For and Access From China

It is better suited to AI security teams, enterprise Agent platform teams, model governance/audit teams, and high-risk scenarios such as finance, credit, cybersecurity, scientific research, and healthcare. Information on access from China, payment, and local support is unknown. For real-world deployment, key points to confirm include network connectivity, whether data leaves China, whether private deployment is supported, and contract-based payment options. Alternative directions include LLM security gateways, Agent monitoring, RAG security testing, model evaluation, and interpretability research tools.