AgentBoundary AI: Secure AI Agent Execution

Created by team The Explorers on November 08, 2026

C1-Blast Radius Zero

AgentBoundary AI is a research prototype developed by The Explorers to address a growing security challenge: how can autonomous AI agents perform useful tasks without receiving unrestricted access to computer systems and sensitive information? Modern AI agents can generate code, interact with tools, and execute commands. However, excessive permissions can lead to unauthorized actions, unexpected code execution, and potential exposure of confidential data. Our solution introduces a task-aware security gateway that separates AI-generated actions from permission decisions. Each task receives limited authorization, and the gateway checks whether permission remains valid before allowing another execution step. Users can revoke permission during a multi-step task, preventing subsequent actions from proceeding. The project also includes a design for disposable Docker sandboxes with restricted network access, filesystem permissions, and resource limits. A containment lab provides controlled scenarios intended to evaluate these restrictions. Our current working prototype is deployed on Render using Python, FastAPI, SQLite, HTML, CSS, and JavaScript. It demonstrates predefined two-step task planning, permission revocation, and audit events. The implementation has passed 27 local automated tests. Vultr Serverless Inference integration and live sandbox execution are planned deployment capabilities that still require validation. Our goal is to make AI-agent workflows more controllable, transparent, and accountable while providing a foundation for future research into secure autonomous execution.

Category tags:

Developer Tools, Security, Automotive, Agent Builder track - The INTERNET OF AGENTS

Github Presentation Demo

Explore more applications

Wayfinder, Lumen OCR and Provenance RAG on ROCm

My three AMD AI League entries are all built on the mandated ROCm PyTorch base image and run Qwen3-VL-8B-Instruct in bf16 on the AMD GPU. A resident server loads the model once and listens on a unix socket, so each harness call to /app/app.py is a thin client that finishes inside the 30-second budget. The weights ship inside the images, and nothing is downloaded at evaluation time. Match 4, Wayfinder (web information retrieval). It reads a natural-language request, fetches the start page and ranks same-site links by how well they match the question. Each page is cleaned to text, with tables kept as rows. The model then either gives the value with the passage it came from, or picks the next link to open, for up to three hops. An answer only counts if it appears on a cited page, and if nothing is found the output is an empty answer. It also works on a saved copy of a site (--corpus). Match 2, Lumen OCR, reads licence plates and road signs. On Chinese plates it keeps the province character; on US plates it drops the state banner and slogan. Multi-line signs are joined in reading order, and when the first pass is unsure about a hard image (glare, blur, noise) it reads an enhanced copy again. Match 3, Provenance RAG, indexes PDFs, Word files, spreadsheets, CSVs, logs, source code and images (read by the vision model) once. Empty folders, unreadable files, encrypted PDFs and unknown types are skipped. Each answer cites only the files it needed, a multi-hop chain cites every file in the chain, and withdrawn revisions are ignored. Each project is also a reusable tool with a CLI, a Python API and a local Studio web UI. On the public samples, using AMD Token Lab's hosted Qwen3.8-27B, Lumen OCR scored 10/10 and Provenance RAG 10/10. I have not yet run the shipped 8B model on an AMD GPU. Match 4 image: docker.io/abuhuraira0001/alx-workbench:m4-v1 Code: github.com/abuhuraira99/wayfinder, github.com/abuhuraira99/lumen-ocr, github.com/abuhuraira99/provenance-rag

Abu Huraira

AMD ROCmQwen3-VLHuggingFace Hub

Quarantine: proof before the install

An agent triages an untrusted package before install: it plans probes, runs them in a hardened container on Vultr, and returns a verdict backed by the egress proxy log, the process result and a filesystem diff the package cannot write.

Kosi IT

DryRun AI

DryRun AI analyzes Python code, executes it inside a restricted Docker sandbox on Vultr, and compares AI-generated predictions with recorded execution evidence in a structured report.

DryRun AI

VULTRAntigravityAnthropic ClaudeChatGPTCodexQwen3

Counterproof: Zero-Blast-Radius Bug Hunter

Find the smallest generated input that breaks your code, then replay it in a fresh sandbox. Vultr Serverless Inference builds the oracle and tests; gVisor isolates execution on a Vultr VM.

Counterproof

Upcoming AI Hackathons
For Innovators & Creators

Explore more applications

Wayfinder, Lumen OCR and Provenance RAG on ROCm

My three AMD AI League entries are all built on the mandated ROCm PyTorch base image and run Qwen3-VL-8B-Instruct in bf16 on the AMD GPU. A resident server loads the model once and listens on a unix socket, so each harness call to /app/app.py is a thin client that finishes inside the 30-second budget. The weights ship inside the images, and nothing is downloaded at evaluation time. Match 4, Wayfinder (web information retrieval). It reads a natural-language request, fetches the start page and ranks same-site links by how well they match the question. Each page is cleaned to text, with tables kept as rows. The model then either gives the value with the passage it came from, or picks the next link to open, for up to three hops. An answer only counts if it appears on a cited page, and if nothing is found the output is an empty answer. It also works on a saved copy of a site (--corpus). Match 2, Lumen OCR, reads licence plates and road signs. On Chinese plates it keeps the province character; on US plates it drops the state banner and slogan. Multi-line signs are joined in reading order, and when the first pass is unsure about a hard image (glare, blur, noise) it reads an enhanced copy again. Match 3, Provenance RAG, indexes PDFs, Word files, spreadsheets, CSVs, logs, source code and images (read by the vision model) once. Empty folders, unreadable files, encrypted PDFs and unknown types are skipped. Each answer cites only the files it needed, a multi-hop chain cites every file in the chain, and withdrawn revisions are ignored. Each project is also a reusable tool with a CLI, a Python API and a local Studio web UI. On the public samples, using AMD Token Lab's hosted Qwen3.8-27B, Lumen OCR scored 10/10 and Provenance RAG 10/10. I have not yet run the shipped 8B model on an AMD GPU. Match 4 image: docker.io/abuhuraira0001/alx-workbench:m4-v1 Code: github.com/abuhuraira99/wayfinder, github.com/abuhuraira99/lumen-ocr, github.com/abuhuraira99/provenance-rag

Abu Huraira

AMD ROCmQwen3-VLHuggingFace Hub

Quarantine: proof before the install

An agent triages an untrusted package before install: it plans probes, runs them in a hardened container on Vultr, and returns a verdict backed by the egress proxy log, the process result and a filesystem diff the package cannot write.

Kosi IT

DryRun AI

DryRun AI analyzes Python code, executes it inside a restricted Docker sandbox on Vultr, and compares AI-generated predictions with recorded execution evidence in a structured report.

DryRun AI

VULTRAntigravityAnthropic ClaudeChatGPTCodexQwen3

Counterproof: Zero-Blast-Radius Bug Hunter

Find the smallest generated input that breaks your code, then replay it in a fresh sandbox. Vultr Serverless Inference builds the oracle and tests; gVisor isolates execution on a Vultr VM.

Counterproof