AI Security Gateway

Appearance

[2026.2.4-beta] ​

🎉 Fourth Public Beta Release ​

The AI Security Gateway is a unified security platform providing real-time monitoring, policy enforcement, and threat detection for Large Language Model (LLM) APIs and Model Context Protocol (MCP) servers. This beta release represents a comprehensive security proxy and monitoring platform for AI infrastructure.

This release introduces the AI Security Skills Hub, a centralised skill approval authority for AI assistants. It also adds Agent-to-Agent (A2A) Card Change Detection to protect against rug-pull attacks, and redesigned dashboards across the platform.

🧩 AI Security Skills Hub ​

Research shows that 12% of skills in public marketplaces contain malicious patterns. The AI Security Skills Hub is a centralised skill approval authority that lets AI assistants submit their skills and plugins for automated security scanning before execution, and gives administrators a web-based approval workflow to decide which skills are safe.

The Skills Hub exposes a full MCP (Model Context Protocol) server that AI assistants like Claude Code, Cursor, and Agent Zero connect to directly. Assistants submit skill source code, check approval status, and report suspicious runtime behaviour, all through standard MCP tool calls.

How it works:

  1. An AI assistant submits skill source code via MCP
  2. The analysis engine scans the code against 43 detection rules across 21 categories with 227 regex patterns, covering shell execution, credential harvesting, reverse shells, obfuscation, prompt injection, crypto wallet theft, and more
  3. Findings are scored by severity (Critical: 50+, High: 30+, Medium: 15+, Low: 1+) and the skill enters the approval queue
  4. An administrator reviews the findings and approves, rejects, or quarantines the skill
  5. AI assistants check approval status before executing any skill

Key features:

  • Content-addressable approval: Approval is tied to the SHA-256 hash of the skill source code. Any code change automatically resets the skill to pending, even if it was previously approved
  • ZIP upload analysis: Administrators can upload .zip files containing skill source code directly from the web UI for per-file analysis with findings attributed to specific files and line numbers
  • Runtime activity reporting: AI assistants can report suspicious behaviour during execution (data exfiltration, credential access, privilege escalation, unexpected network activity, behaviour changes), which automatically creates security alerts
  • System prompt template: A ready-to-use prompt template that instructs AI assistants to check skills on startup, submit new skills for analysis, and default to strict mode if the gateway is unreachable
  • Full admin dashboard: Stats, risk distribution, submission history, registry management, and activity reports all in one view

🔄 A2A Card Change Detection ​

A new security monitoring feature that continuously watches for rug-pull attacks against registered Agent-to-Agent (A2A) agents. A rug-pull attack occurs when a remote agent silently modifies its AgentCard after registration, potentially redirecting traffic, injecting malicious skills, or altering its capabilities without the administrator's knowledge.

The card monitor runs on a configurable schedule (default: every 6 hours), re-fetches each agent's card from its remote endpoint, and compares it field-by-field against a stored baseline snapshot.

Detected change types:

  • URL Changed (Critical): Endpoint URL redirected, potential traffic hijacking
  • Name Changed (High): Agent identity was modified
  • Protocol Version Changed (High): A2A protocol version altered
  • Skill Removed (High): A previously available skill was silently dropped
  • Description Changed (Dynamic): Severity scales with similarity; a near-complete rewrite scores Critical, minor edits score Low
  • Skill Added (Medium): A new skill appeared that wasn't present at registration
  • Capabilities Changed (Medium): Streaming, push notification, or state flags toggled
  • Version Changed (Low/Medium): Version string updated; major version bumps score Medium
  • Skill Description Changed (Dynamic): Per-skill description changes scored by text similarity
  • Skill Tags Changed (Low): Skill metadata tags modified

Changes automatically generate security alerts with full metadata and are visible in a dedicated Card Change Monitor tab in the A2A dashboard, where administrators can view change history, inspect old vs. new values, and manage baselines

🎨 Redesigned Dashboards ​

Dashboards across the platform have been redesigned for improved clarity and usability, with cleaner layouts, better data density, and more consistent styling.