Chatbots & APIs
Explore prompt injection and information disclosure through browser chatbots or supported API connections. Keep each probe with the response it produced.
ResearchNode tools · Product preview
A local workbench for assessing chatbots, retrieval systems, AI agents and MCP connections. Plan a test, follow what happened and review the evidence in one project.

The images on this page show the supplied interface preview and local teaching sample. The example finding illustrates the workflow; it is not a result from a customer system.
The tool
PokeAI is the desktop experience built on the AdverScope assessment framework. It connects test configuration, recorded interactions and human review.
Explore prompt injection and information disclosure through browser chatbots or supported API connections. Keep each probe with the response it produced.
Assess how retrieved documents and other external content influence a system. Compare test behavior with legitimate tasks and inspect the supporting records.
Examine tool access, permissions and multi-step actions using configured targets and evidence checks. Available coverage depends on the target and its integrations.
01 / Set up
Create a new assessment or reopen an existing one. Project records, settings and evidence stay organized together in your chosen workspace.
Guided setup connects your planning model and target, records permission and limits, and lets you review the plan before a run. A built-in sample introduces the workflow without connecting a live target.

02 / Test and review
A result is the beginning of the review. Follow the conversation, inspect the retained evidence and decide what needs action.
The results view separates items needing review, accepted issues and uncertain checks. It shows the run status and recorded target-action usage alongside the next step.

The conversation view pairs each test message with the target response. Retained text and repeat tests help a reviewer understand what the system actually did.

Review the evidence before accepting a finding or exporting a report. An uncertain or untested check does not establish that a system is secure. PokeAI remains a beta tool for controlled, authorized assessments.

03 / Work your way
The desktop app offers light, dark and system appearance, with the same project history, findings and scope in each view.
Project records are stored locally. Model-assisted work can use a local model or an approved remote provider; the selected connection determines where model requests are processed.
Continue exploring
Pair the workbench with ResearchNode’s AI security explanations and Project 2501’s hands-on exercises.