AIShield Security Scanner

Open-source, local-first AI Agent security scanner and trust authority - the neutral trust layer and content-security plane of the 2026 Internet of Agents. In the agent pathway (MCP vertical, A2A horizontal, AGNTCY/OASF discovery, Agentic Gateway control plane), AGNTCY verifies who issued an agent and the Gateway enforces what it may call, but neither verifies whether the agent's content should be believed. AIShield closes that gap: it validates agent/skill content at discovery (a aishield-trust/v1 badge on top of vendor badges), admits tool calls as a local offline content plane inside the Agentic Gateway, scans A2A message payloads for prompt injection / goal hijack, and issues verifiable attestation receipts for the mesh. Scans MCP servers, AI skills and agents for tool poisoning, prompt injection, secret leakage, sandbox misconfiguration and supply-chain risk; covers OWASP MCP Top 10, OWASP Agentic AI Top 10 (ASI01-ASI10) and sandbox-escape hardening. Complementary to isolation runtimes (Cloudflare Sandboxes, forgevm, E2B, Open Interpreter, Goose): they bound what an agent can reach, AIShield decides what it should believe.

Skills

• Security Scan
  Scan an MCP server, AI skill or agent description against 264 MCP / 291 skill rule categories (OWASP MCP Top 10 + Agentic AI Top 10 + sandbox hardening). For skill assets, Markdown is treated as executable payload rather than documentation.
  securityauditmcpagent
• Agentic AI Audit
  Audit an AI agent against OWASP Agentic AI Top 10 (ASI01-ASI10): goal hijack, tool misuse, identity abuse, supply chain, code execution, memory poisoning, inter-agent comms, cascading failure, human-agent trust, rogue agents.
  agenticowaspaudit
• Supply Chain & Hallucinated Package Audit
  Offline detection of slopsquatting / AI-hallucinated dependencies in package.json, requirements.txt and pyproject.toml. Covers typosquat (Levenshtein), homoglyph poisoning, brand impersonation, composite hallucination (the ~50% of fabricated names that are NOT edit-distance-similar to any real package, e.g. react-codeshift), cross-registry confusion, dependency confusion, install-script poisoning, untrusted sources, unpinned versions and missing lockfiles. Zero network calls, zero package database.
  supply-chainslopsquattingtyposquatsbomoffline
• Multi-Client MCP Config Discovery & Audit
  Auto-discover MCP server configurations across 14 client surfaces (Claude Desktop, Claude Code user+project, Cursor user+project, VS Code user+project, Windsurf, Gemini CLI, GitHub Copilot CLI, Augment, Zed, Cline, WorkBuddy) and statically audit them for privileged launch, runtime package fetch at startup, shell-interpreter invocation, non-registry provenance, inline plaintext credentials, insecure transport, wildcard bind, unauthenticated remote endpoints, project-level trust traps, namespace shadowing between servers, and 7 classes of toxic capability flows. PURELY STATIC: AIShield never executes any command defined in a scanned configuration - unlike scanners that spawn the server process to read tools/list.
  mcpconfigdiscoverystatic-analysisnamespace-shadowingtoxic-flowlocal-first
• Agent Computer Pre-Flight Scan
  Scan an agent workspace BEFORE the sandbox boots. Parses .mcp.json, forge / agent-forge, Goose and Open Interpreter configurations plus every skill file, scores each item, and returns a boot / review / refuse verdict. Complements isolation runtimes (Cloudflare Sandboxes and Containers, forgevm, E2B, Open Interpreter, Goose) which bound blast radius but do not inspect the content an agent loads inside the box. Also checks 11 sandbox-hardening rules on the box definition itself: mounted docker.sock, --privileged, host network/PID/IPC namespaces, cap_add ALL, CAP_SYS_ADMIN, seccomp=unconfined, --user 0, Kubernetes hostPath. PURELY STATIC: never spawns a command found in the workspace, never fetches the network.
  agent-computersandboxpreflightworkspacestatic-analysislocal-firstcloudflare-sandboxforgevmgooseopen-interpreter
• Continuous Attestation
  Subscribe an MCP server, skill or live agent workspace to recurring re-scanning (default 7-day cycle). Detects drift against the recorded evidence hash, revokes certification when the score drops below threshold, and exposes a machine-readable answer to 'is this still trustworthy right now'. Designed for rug-pull defence: certification without expiry is marketing.
  attestationcertificationrug-pullmonitoringtrust
• Trust Score Lookup
  Return an agent's AIShield Trust Score (0-100) and certification level from the Agent Registry.
  trustregistryscore
• Agent Identity & Credential Scan
  Scan the agent identity layer (NHI). Verifies AgentCard / agent-identity declarations are signed (JWS/DID/proof), credentials are short-lived rather than never-expiring, authorization is least-privilege (flags scope:'*' and over-broad grants that violate scope attenuation), and mTLS/DID verification is present. This is the fastest-moving front of 2026 agent security (the top A2A issues are all identity; Authentik's NHI wave; ANS/DNSid/Entra Agent ID). AIShield both issues trust certificates AND audits identity defects.
  identitynhiagent-cardscope-attenuationmtlsdida2a
• Agent Network / Mesh Config Scan
  Scan the agent network layer. Flags Cloudflare Mesh / VPC bindings that expose the whole account network to every agent (the gap Cloudflare itself admits: 'per-agent identity and policy evaluation are future work'), unauthenticated agent endpoints (auth: none), bind-to-all-interfaces exposure (0.0.0.0), and private/internal resources marked public:true. Answers the 'trust shallow' problem left open by A2A's signed AgentCard: content trust + identity attribution + network reachability.
  networkmeshcloudflare-meshvpcreachabilityexposure
• Attack Replay & Regression Detection
  Snapshot and replay past attack payloads against the current rule set. Detects rule-regression: a payload that was previously blocked but is now allowed because rules were weakened or a pattern was missed. Each snapshot stores payload hash + verdict + evidence, enabling 'has our defense regressed since last check' audits. Borrowed from the ChronosFix 'fault time machine' pattern in agent infra competitions.
  attack-replayregressionchronos-fixsnapshotdefense-hardening
• Vertical-Domain Semantic Risk Scan
  Domain-specific semantic risk screening for high-sensitivity verticals: finance (fraud inducement / unlicensed wealth management / pump-and-dump), medical (unlicensed diagnosis / false cure claims), and government/public-sector (sensitive topics / unauthorized disclosure). Sits on top of the generic OWASP rule set to catch agent output that is technically compliant but semantically dangerous in its context. Borrowed from the FinFlux 'financial semantic admission' pattern.
  vertical-risksemantic-admissionfinancemedicalgovfinflux

How to call

https://aishield.tools/api/v1/mcp

Listed in

Directories this entry was found in.