AgentTrust
Quality verification for AI agents and MCP servers. 6-axis scoring, adversarial probes.
0Tools
23Findings
0Stars
Mar 19, 2026Last Scanned
9 critical · 12 high · 1 medium · 1 low findings detected
Security Category Deep Dive
Prompt Injection
Prompt & context manipulation attacks
69
Maturity
14
Rules
5
Sub-Categories
1
Gaps
64%
Implemented
56
Tests
1
Stories
100%3 rules
Injection via tool descriptions and parameter fields
GAP-001Prompt Injection Coverage GapMissing detection coverage for emerging prompt injection attack variants not addressed by current rules
100%4 rules
Hidden instructions via external content and tool responses
100%2 rules
Context window saturation and prior-approval exploitation
100%3 rules
Payload hiding via invisible chars, base64, schema fields
100%2 rules
Injection via prompt templates and runtime tool output
Findings23
9critical
12high
1medium
1low
Critical9
criticalQ9Agentic Workflow DAG ManipulationMCP05-privilege-escalationAML.T0058
Pattern "(?:next[_\s]?(?:step|node|agent)|route[_\s]?to|redirect[_\s]?to|skip[_\s]?(?:step|node|validation))" matched in source_code: "route to" (at position 44300)
MCP tools must NOT modify agentic workflow graphs, execution order, or routing during runtime. Workflow structure should be immutable once execution begins. If dynamic workflow modification is required, implement: (1) approval gates that require human confirmation before graph mutations, (2) integrity checks that validate the workflow graph against a known-good baseline after each step, (3) audit logging of all graph modifications with rollback capability. Reference: arXiv 2602.19555, Trend Micro subgraph impersonation attack (2026).
criticalK14Agent Credential Propagation via Shared StateMCP05-privilege-escalationAML.T0054
Pattern "(process\.env|os\.environ|setenv|putenv).*(?:token|credential|api[_\s-]?key|secret|password)" matched in source_code: "process.env.AGENTTRUST_API_KEY" (at position 3764)
Never write credentials to shared agent state. Use credential vaults (HashiCorp Vault, AWS Secrets Manager) with per-agent scoped access. Implement OAuth token exchange (RFC 8693) for cross-agent authorization. Redact credentials from all agent outputs before writing to shared memory. Required by OWASP ASI03/ASI07 and MAESTRO L7.
criticalK8Cross-Boundary Credential SharingMCP05-privilege-escalationAML.T0054
Pattern "(return|respond|output|result).*(?:token|credential|api[_\s-]?key|secret|password|bearer)" matched in source_code: "return { apiKey" (at position 4090)
Never forward, share, or embed credentials across trust boundaries. Use OAuth token exchange (RFC 8693) to create scoped, delegated tokens instead of passing original credentials. Never include credentials in tool responses. Required by ISO 27001 A.5.17 and OWASP ASI03.
criticalL9CI/CD Secret Exfiltration PatternsMCP07-insecure-configAML.T0057
Pattern "(?:btoa|Buffer\.from|base64\.b64encode).*(?:process\.env|os\.environ|TOKEN|SECRET|KEY|PASSWORD)" matched in source_code: "Buffer.from(publicKey" (at position 7841)
Never print, log, or transmit CI environment variables containing secrets. Use GitHub Actions '::add-mask::' to prevent accidental secret exposure in logs. Audit all third-party Actions for secret access patterns. Use OIDC tokens instead of long-lived secrets where possible. Restrict secret access to specific workflow jobs and steps. Monitor CI logs for base64-encoded strings.
criticalQ11Code Suggestion Poisoning via MCPMCP01-prompt-injectionAML.T0054.001
Pattern "(?:suggest|generate|complete|insert).*(?:code|function|class|import|require)" matched in source_code: "completed", "input-require" (at position 22215)
MCP tool outputs flowing into IDE code suggestion contexts must be sanitized. Implement output content policies that: (1) strip hidden Unicode characters (zero-width, RTL override, tag characters), (2) detect embedded instructions targeting AI code assistants, (3) validate code blocks against security patterns before they enter the suggestion pipeline, (4) never include shell commands in tool outputs without explicit [COMMAND] markers visible to the user. Reference: IDEsaster (Dec 2025), arXiv 2509.22040.
criticalQ13MCP Bridge Package Supply Chain AttackMCP10-supply-chainAML.T0054
Pattern "["']@modelcontextprotocol/sdk["']\s*:\s*["'](?:\^|~|\*|latest)" matched in source_code: ""@modelcontextprotocol/sdk": "^" (at position 47576)
MCP bridge packages (mcp-remote, mcp-proxy, @modelcontextprotocol/sdk, fastmcp) are high-value supply chain targets — CVE-2025-6514 (CVSS 9.6) in mcp-remote affected 437,000+ installs. Always pin exact versions (no ^ or ~ ranges). Use lockfiles (package-lock.json, pnpm-lock.yaml, uv.lock). Never run `npx mcp-remote` without version pinning. Verify package integrity with `npm audit` or `pip-audit` before deployment. Reference: CVE-2025-6514, OWASP ASI04.
criticalQ3Localhost MCP Service HijackingMCP07-insecure-configT1557
Pattern "(?:express|fastify|koa|hono).*(?:listen|server)(?!.*(?:host[_\s]?(?:check|valid|verify)|helmet))" matched in source_code: "express.json() globally — StreamableHTTPServer" (at position 43862)
MCP servers binding to localhost must: (1) validate the Host header to prevent DNS rebinding attacks (CVE-2025-49596), (2) set strict CORS origins instead of wildcard '*', (3) require authentication tokens even for local connections, (4) use random high ports instead of predictable defaults. For stdio transport, validate all input at the JSON-RPC level before processing. Consider using Docker MCP Gateway or similar container isolation.
criticalQ6Agent Identity Impersonation via MCPMCP05-privilege-escalationAML.T0054
Pattern "(?:agent[_\s-]?id|agent[_\s-]?name|agent[_\s-]?role|caller[_\s-]?agent|source[_\s-]?agent).*(?:param|arg|input|string)" matched in source_code: "agentId: string" (at position 1258)
MCP tools in multi-agent systems must verify agent identity cryptographically — never accept agent_id/agent_role as plain string parameters. Use cryptographic attestation (signed tokens, mTLS certificates, or capability tokens) for inter-agent communication. Implement the principle of least privilege: each agent should only be able to claim its own identity. Reference: OWASP ASI03, arXiv 2602.19555.
criticalC1Command InjectionMCP03-command-injectionAML.T0054
Pattern "`[^`]+`" matched in source_code: "`${base}/`" (at position 2827)
Replace exec()/execSync() with execFile() and pass arguments as an array, never as a string. Validate all inputs against an allowlist before use in any shell context. For subprocess.run, always pass a list and shell=False.
High12
highK15Multi-Agent Collusion PreconditionsMCP05-privilege-escalationAML.T0054
Pattern "(agent|delegate).*(?:send|receive|message|communicate)(?!.*(?:log|audit|trace|record|monitor))" matched in source_code: "agenttrust_send" (at position 21149)
Implement collusion-resistant multi-agent architecture: (1) Verify agent identity cryptographically before accepting commands, (2) Apply ACLs to shared write surfaces, (3) Rate-limit cross-agent invocations, (4) Audit all inter-agent communication with timestamps and agent IDs, (5) Baseline normal interaction patterns for anomaly detection. Required by MAESTRO L7 and CoSAI MCP-T9.
highQ14Concurrent MCP Server Race ConditionMCP07-insecure-configT1068
Pattern "(?:read|write|modify|delete).*(?:file|path|directory)(?!.*(?:lock|mutex|semaphore|flock|atomic))" matched in source_code: "readConfigFile" (at position 2999)
MCP servers sharing filesystem or database backends with other servers must implement proper concurrency controls. Use: (1) file locking (flock/lockfile) for filesystem operations, (2) database transactions for all read-modify-write sequences, (3) atomic file operations (O_EXCL, mkdtemp) instead of check-then-create, (4) lstat() to detect symlinks before following (CVE-2025-53109). Never assume exclusive access to shared resources — other MCP servers may be operating concurrently.
highQ15A2A/MCP Protocol Boundary ConfusionMCP06-excessive-permissionsAML.T0054
Pattern "(?:mcp|model[_\s-]?context).*(?:a2a|agent[_\s-]?to[_\s-]?agent)" matched in source_code: "MCP server for AgentTrust — A2A" (at position 46628)
Servers bridging A2A and MCP protocols must: (1) sanitize all A2A task metadata before passing to MCP tool inputs, (2) apply MCP content policies to A2A TextPart/FilePart/DataPart content, (3) validate A2A push notifications before they re-enter MCP context, (4) require cryptographic verification for agent discovery and registration (prevent fake agent advertisement — arXiv 2602.19555), (5) maintain separate permission models for A2A and MCP operations — trust in one protocol must not automatically grant trust in the other.
highN10Incomplete Handshake Denial of ServiceMCP07-insecure-configAML.T0054
Pattern "(?:createServer|listen)\s*\((?!.*(?:maxConnections|maxClients|connectionLimit|MAX_CONN))" matched in source_code: "createServer(" (at position 19053)
Enforce a handshake timeout (recommended: 30 seconds) — terminate connections that do not complete the initialize handshake within the deadline. Limit maximum concurrent pending connections. An attacker can exhaust server connection slots by initiating MCP connections without completing the handshake (Slowloris-style attack). Reference: MCP spec 2025-03-26 lifecycle — initialize MUST complete before functional requests.
highO10Privacy-Violating TelemetryMCP04-data-exfiltrationAML.T0057
Pattern "(?:os\.(?:type|release|platform|arch|hostname|cpus|networkInterfaces|userInfo|homedir)|process\.(?:arch|platform|ppid))\s*\(?" matched in source_code: "os.homedir(" (at position 1026)
Minimize data collection to only what is strictly necessary for tool functionality. Never collect OS fingerprints, hardware IDs, network interface details, or unique device identifiers without explicit user consent. Remove all tracking pixels, analytics beacons, and behavioral profiling code. If telemetry is needed, make it opt-in only, clearly documented, and limited to non-identifying aggregated metrics. Comply with GDPR data minimization (Art. 5(1)(c)) and EU AI Act transparency requirements (Art. 52).
highO6Server Fingerprinting via Error ResponsesMCP04-data-exfiltrationAML.T0057
Pattern "(?:res\.(?:send|json)|return|response).*(?:__dirname|__filename|process\.cwd\(\)|path\.resolve|path\.join)" matched in source_code: "return path.join" (at position 7449)
Never expose process, OS, runtime, or database metadata in tool responses or error messages. Use generic error messages ("An error occurred") for production responses. Remove or disable debug/diagnostic endpoints. If health endpoints are needed, limit them to simple "ok"/"error" status without infrastructure details. Wrap all error handlers with a sanitization layer that strips system information.
highC3Server-Side Request Forgery (SSRF)MCP04-data-exfiltrationAML.T0057
Pattern "\bfetch\s*\(\s*(?:req|request|input|param|params|args|url|uri|href|link|target|destination|endpoint|host|address|resource|src|source|location|proxy|callback|userUrl|user_url|externalUrl|external_url|redirectUrl|redirect_url)" matched in source_code: "fetch(url" (at position 6846)
Validate ALL user-supplied URLs before making HTTP requests:
1. Parse the URL and check the hostname against an explicit allowlist of permitted domains.
2. Block requests to RFC 1918 private ranges: 10.0.0.0/8, 172.16.0.0/12, 192.168.0.0/16.
3. Block loopback (127.0.0.0/8), link-local (169.254.0.0/16), and IPv6 equivalents.
4. Block file:// and other non-http(s) protocols explicitly.
5. Disable automatic redirect following, or re-validate each redirect destination.
6. In cloud environments: block requests to IMDS endpoints (169.254.169.254,
metadata.google.internal) at both the application AND network layer.
Example (Node.js): Use the `ssrf-req-filter` package or implement URL validation
against an allowlist before calling fetch/axios/got.
highK16Unbounded Recursion / Missing Depth LimitsMCP07-insecure-configAML.T0054
Pattern "function\s+(\w+).*\{[^}]*\1\s*\((?!.*(?:depth|level|limit|max|count|recursi))" matched in source_code: "function loadRuntimeConfig(opts?: { refresh?: boolean }): Promise<RuntimeConfig> {
const refresh = opts?.refresh === true;
if (!refresh && runtimeConfigCache) return runtimeConfigCache;
if (!refresh && runtimeConfigPending) return runtimeConfigPending;
runtimeConfigPending = loadRuntimeConfigInternal(" (at position 5381)
Add explicit depth/recursion limits to all recursive operations. Use iterative approaches where possible. Set maximum depth for directory walking (max_depth=10), tree traversal (max_level=20), and agent re-invocation (max_calls=5). Implement circuit breakers that halt after N iterations. Required by EU AI Act Art. 15 (robustness) and OWASP ASI08.
highI15Transport Session SecurityMCP07-insecure-configAML.T0054
Pattern "session[_\s-]?(token|key|id)\s*[:=]\s*(req\.|params\.|query\.|body\.)" matched in source_code: "sessionId = req." (at position 44208)
Use HTTPS for all MCP Streamable HTTP endpoints. Generate cryptographically random session IDs (min 128 bits entropy). Do not accept session IDs from user input (CVE-2025-6515). Validate TLS certificates — do not disable certificate verification.
highK11Missing Server Integrity VerificationMCP10-supply-chainAML.T0054
Pattern "(connect|load|register|add)[_\s-]?(mcp|server|tool)(?!.*(?:verify|validate|checksum|hash|sign|cert|fingerprint|pin))" matched in source_code: "connect server" (at position 44778)
Implement cryptographic verification for MCP server connections: (1) Pin server TLS certificates or public keys, (2) Verify server tool definition checksums against a known-good manifest, (3) Use package manager integrity checks (npm integrity, pip --require-hashes). The MCP spec recommends but doesn't yet mandate server signing — implement it proactively. Required by ISO 27001 A.8.24 and CoSAI MCP-T6.
highK13Unsanitized Tool OutputMCP02-tool-poisoningAML.T0054
Pattern "(?:query|execute|select|find).*(?:return|respond|result|rows|data)(?!.*(?:sanitize|escape|encode|map|filter|select|pick))" matched in source_code: "query to filter result" (at position 26001)
Sanitize all external data before including in tool responses. Implement output encoding that neutralizes prompt injection patterns. Truncate excessively long content. Validate structure before passing database results. Apply the principle: treat all external data as untrusted, even in tool outputs. Required by CoSAI MCP-T4.
highO8Timing-Based Covert ChannelMCP04-data-exfiltrationAML.T0057
Pattern "(?:sleep|delay|timeout|setTimeout|wait).*(?:>>|<<|\&\s*(?:1|0x[0-9a-f]+))" matched in source_code: "wait apiRequest<Record<string, unknown>>" (at position 13835)
Remove all code that calculates sleep/delay durations from application data, secrets, or any variable-length content. Tool response times should be constant or determined only by legitimate processing time. If rate limiting is needed, use fixed intervals not derived from data values. Monitor for anomalous response time patterns that could indicate timing-based exfiltration.
Medium1
mediumK17Missing Timeout or Circuit BreakerMCP07-insecure-configAML.T0054
Pattern "(?:fetch|axios|got|request|urllib|httpx|http\.get|http\.post)\s*\((?!.*(?:timeout|signal|AbortSignal|deadline|cancel))" matched in source_code: "fetch(" (at position 6846)
Add timeouts to ALL external calls: HTTP requests (30s), database queries (10s), subprocess execution (60s), and MCP tool calls (30s). Implement circuit breakers that open after N consecutive failures (e.g., opossum, cockatiel). Use AbortSignal for cancellable operations. Required by EU AI Act Art. 15 and OWASP ASI08.
Low1
lowF4MCP Spec Non-ComplianceMCP07-insecure-config
Server fails MCP spec compliance checks: required:server_name; required:server_version; required:protocol_version; recommended:tool_descriptions; recommended:parameter_descriptions
Follow the MCP specification for server metadata. Include server name, version, and protocol version. Provide descriptions for all tools and parameters.