.png&w=256&q=75)
🤓 Latest Submissions
AgentGuard
As AI coding agents become capable of connecting to MCP servers and external tools, developers can move from idea to implementation faster than ever. But that speed creates a difficult trust problem: a tool that was safe when approved can later change its capabilities, permissions, schemas, or dependencies without everyone noticing. Developers need to move quickly, while security teams need confidence that what was approved is still what is running. AgentGuard is a security control layer for AI-assisted development and agent toolchains. It starts with MCP and answers a simple question: **Can we trust this agent tool, and is it still the same tool we approved?** AgentGuard discovers MCP tools, normalizes their capabilities, creates a versioned Capability Passport, and evaluates them against deterministic security policies. When a tool changes, AgentGuard compares the approved state with the current state and highlights added, removed, or modified capabilities. Risky or unauthorized changes can be blocked or sent for review, while the system preserves an evidence trail of what changed, why it mattered, and how the decision was reached. The primary users are developers, platform and DevEx teams, AppSec/security teams, and organizations adopting AI agents that need stronger control over their toolchain. A developer can connect a tool, review its Capability Passport, approve a known-good version, and continue working normally. If an AI-assisted change later expands the tool's permissions or capabilities, AgentGuard detects the drift and stops the change from silently becoming trusted. What makes AgentGuard different is its focus on **capability integrity** rather than simply detecting generic threats. Trust is attached to a specific, versioned capability state. A meaningful change triggers re-evaluation instead of inheriting previous approval. This creates a clear chain from tool identity and capability state to policy decision and evidence. AgentGuard is designed to add a security checkpoint without forcing developers to give up the speed that makes AI-assisted development valuable: **build quickly, verify continuously, and never trust a tool simply because it was trusted yesterday.**
27 Sep 2026
.png&w=828&q=75)
TABLEMIND — Bimanual VLA
TABLEMIND is a simulation-first Physical AI system for bimanual manipulation using two simulated SO-101 arms in MuJoCo. It interprets natural-language instructions together with simulated camera observations, maintains task context, plans multi-step actions, coordinates both manipulators, and verifies the resulting scene state. The system is designed around Hugging Face LeRobot and SmolVLA-compatible policy workflows, with a task-planning layer for shared-workspace reasoning, collision-aware sequencing, object hand-offs, complementary dual-arm actions, and recovery when the scene changes. TABLEMIND also includes environment randomization and evaluation tooling to measure robustness across different object placements and physical and visual conditions. For deployment optimization, the project integrates the Intel OpenVINO and OpenVINO Physical AI inference path, with support for model conversion, optimization, precision selection, device benchmarking, and Intel CPU/iGPU/NPU execution where supported. The final target is an Intel Core Ultra Series 2/3 system, while the robotic environment remains fully simulation-first as required by the online challenge. The project is built to be reproducible, measurable, and demonstrable through randomized MuJoCo evaluations, inference benchmarks, and a clear end-to-end perception-to-action workflow.
16 Sep 2026
