$1.5B stolen in single Bybit hack -- North Korea's Lazarus Group CBS News///8,000+ MCP servers exposed with ZERO authentication Trend Micro///Pentagon designates Anthropic a "supply chain risk" CBS News 2026///5 of top 7 most-downloaded AI agent skills were MALWARE Snyk///McKinsey breached through exposed AI agent APIs Security Boulevard///47 enterprise deployments compromised via AI plugin ecosystem HackerNoon///Claude Code RCE vulnerability -- execute code via poisoned repos Check Point///China BANS OpenClaw from all government agencies TechWire Asia///1 BILLION records stolen from 39 companies in single breach Bright Defense///48% of security pros: agentic AI is THE most dangerous attack vector Dark Reading///
ddot Labs -- Agentic AI Industry Security Assessment -- March 2026

1,693 MCP Servers.
15 Million Lines of Code.
The Industry Is Not Ready.

The largest independent security assessment of the Model Context Protocol ecosystem ever conducted. We analyzed every publicly available MCP server. The results demand immediate action. This is the first in a series of ongoing MCP ecosystem security assessments.

1,693
Servers Analyzed
44%
Failing (Grade F)
59.8
Average Score /100
9,138
Security Findings
1.4%
Enterprise Ready

Executive Summary

ddot Labs conducted a comprehensive automated security assessment of 1,693 MCP servers -- representing effectively the entire public MCP ecosystem. The analysis covered 15.2 million lines of code across TypeScript, Python, Go, Rust, Java, C#, and Ruby implementations.

746
Servers Failing (F)
539
Servers at D Grade
24
Enterprise Ready (A)
212
Critical Findings
1,175
High-Severity Findings
85%
No Audit Logging
73%
No Rate Limiting
64%
No Sandboxing
53%
No Authentication
41%
Command Execution
19%
SQL Injection Risk
7%
Hardcoded Secrets

"Every enterprise deploying MCP servers today is operating without the security controls they would demand from any other infrastructure software. The gap between what organizations assume they have and what actually exists is not a crack -- it is a chasm."

-- ddot Labs Security Research, March 2026
746
Servers Failing Security
44% of all MCP servers score below 60/100, failing to meet even basic security standards. These servers are deployed in production environments across enterprises globally.
693
Command Execution Vectors
41% of servers contain direct OS command execution (child_process, subprocess, exec). Combined with absent input validation, these are remote code execution vulnerabilities.
329
SQL Injection Risks
19% of servers construct SQL queries through string concatenation instead of parameterized queries. Every database-connected server is a potential data breach vector.
121
Hardcoded Secrets in Code
7% of servers contain API keys, passwords, or secrets hardcoded directly in source code. These credentials are publicly visible on GitHub right now.

Scope & Methodology

This assessment represents the most comprehensive security analysis of the MCP ecosystem ever conducted, covering servers across all major programming languages and use cases.

Scope

Servers Analyzed1,693
Lines of Code15,227,815
Source Files~200,000+
LanguagesTypeScript, Python, Go, Rust, JavaScript, Java, C#, Ruby
SourcePublic GitHub repositories from Awesome MCP Servers directory + official Anthropic repos
DateMarch 24, 2026

Analysis Engine

Detection Patterns30+
Security Categories13
Standards MappedNIST 800-53, OWASP LLM Top 10, CMMC Level 1
MethodStatic pattern analysis + architecture assessment + supply chain review
ExecutionParallel analysis across 32 CPU cores, all source processed locally
ReproducibilityFully automated, deterministic, auditable

Language Distribution

750
TypeScript
633
Python
119
Go
99
JavaScript
51
Rust
41
Java/C#/Ruby

Grade Distribution

The bell curve skews heavily toward failure. 76% of all MCP servers score D or below. Only 1.4% achieve enterprise readiness.

746
F
44.1%
539
D
31.8%
283
C
16.7%
101
B
6.0%
24
A
1.4%
75.9% of all MCP servers (1,285 out of 1,693) score D or F, failing to meet minimum security standards for production deployment. These servers are deployed across enterprises operating in healthcare, finance, government, and critical infrastructure today.

Finding Prevalence

How common each security gap is across all 1,693 servers. These are not edge cases -- they are the norm.

MED No Audit Logging
85%
MED No Rate Limiting
73%
MED No Sandboxing
64%
MED No Output Sanitization
61%
MED No Authentication
53%
LOW No Dependency Lockfile
47%
HIGH Command Execution
41%
MED No Input Validation
29%
HIGH SQL Injection Risk
19%
MED No Encryption at Rest
19%
MED Unrestricted Network
16%
CRIT Hardcoded Secrets
7.1%
HIGH Dynamic Code Execution
4.7%
CRIT Hardcoded API Keys
3.8%
HIGH Path Traversal
2.1%
HIGH Unsafe Deserialization
1.8%
CRIT AWS Credentials Exposed
1.5%

Security by Category

Enterprise-critical categories -- Database, Cloud, and Memory/Knowledge servers -- score the lowest. The servers handling the most sensitive data have the weakest protections.

Category Servers Avg Score Score Risk Assessment
Memory / Knowledge 33 51.5
CRITICAL Stores user data, conversation history, knowledge graphs -- with the lowest avg score
Media 18 51.6
HIGH Image/video/audio processing with command execution risks
Database 42 52.5
CRITICAL Direct production database access with SQL injection and no auth
Data / Analytics 32 54.2
HIGH Data pipeline access without encryption or access control
Filesystem 10 55.8
CRITICAL Direct host filesystem read/write with path traversal risk
AI / LLM 46 56.8
HIGH LLM integrations without prompt injection protection or output filtering
Browser / Web 32 57.6
HIGH Browser automation with unrestricted navigation and command execution
Cloud / Infra 31 58.3
CRITICAL AWS/Azure/GCP access with hardcoded credentials and no isolation
Security 35 60.1
HIGH Security-focused servers that fail their own security standards
General 1,213 60.3
MEDIUM Broad category, systemic lack of security controls
DevTools 31 60.5
MEDIUM Developer tooling with elevated privileges and command execution
Communication 26 61.4
MEDIUM Slack/Discord/email access without authentication controls
Git / Code 74 62.1
MEDIUM Source code access with command execution vectors
Search 40 64.0
MEDIUM External search APIs without result sanitization
CRM / Business 15 70.7
MEDIUM Best category score but still below enterprise threshold
The 5 categories handling the most sensitive enterprise data -- Database, Memory/Knowledge, Filesystem, Cloud/Infrastructure, and Data/Analytics -- all score BELOW 60/100. The servers with the highest blast radius have the weakest defenses.

Critical Failures & Ironies

Some findings go beyond concerning into absurd. Security-branded servers that fail security. Secrets managers that leak secrets. These examples illustrate the depth of the industry's problem.

Servers Scoring Zero

Eight servers received the minimum possible score of 0/100, meaning every security check failed.

empathy-framework
Smart-AI-Memory | Python
0
Critical: hardcoded API keys, secrets, AWS creds | High: cmd exec, eval, path traversal
skylos
duriantaco | Python
0
Critical: 3 secret types exposed | High: 4 injection vectors including SQL
open-code-review
raye-deng | TypeScript
0
Critical: 3 secret types | High: cmd exec, eval, SQL injection, path traversal
mcp-adr-analysis-server
tosin2013 | TypeScript
0
Critical: 3 secret categories | High: cmd exec, eval, SQL injection, unsanitized paths

The Ironies

When the product's name promises security but the code delivers the opposite.

Pantheon-Security / notebooklm-mcp-secure
5
A server with both "Security" in the organization name AND "secure" in the project name scores 5/100. Contains hardcoded API keys, hardcoded secrets, and AWS credentials in source code. The word "secure" appears in its name. Its score is 5.
forest6511 / secretctl
17
A secrets management tool that contains hardcoded secrets in its own source code. The tool designed to protect secrets cannot protect its own.
Security Category Average
60
35 servers in the "Security" category -- tools explicitly designed for security purposes -- average a score of 60.1/100. Nearly half (15/35) are outright failing. The security tools are not secure.

Enterprise-Critical Failures

Servers deployed in enterprise environments with critical access to sensitive systems.

Category Servers Failing (F) Failure Rate Risk
Database Servers 42 24 57% Direct access to production data with SQL injection and no auth
Cloud/Infrastructure 31 15 48% AWS/Azure/GCP control plane access, credential exposure
Security Tools 35 15 43% Security tools that fail security assessment
Finance 15 6 40% Payment processing, trading, crypto operations
Communication 26 9 35% Slack, email, messaging -- corporate communications

The Solution: ddot Security Gateway

ddot addresses every finding in this report. Every red bar above turns green when agents operate through the ddot security gateway.

Security Control Industry (1,693 Servers) ddot Gateway
Sandboxing ABSENT in 64% -- Full process privileges Wasm sandbox (Wasmtime) -- DENY-ALL capabilities, fuel metering, memory ceilings
Authentication ABSENT in 53% -- Anonymous tool invocation Ed25519 signing chain (Master -> CA -> Signing Key) with CRL
Audit Trail ABSENT in 85% -- Zero forensic trail Tamper-evident SHA-256 chain with Bitcoin OP_RETURN attestation
Rate Limiting ABSENT in 73% -- Unlimited invocations Per-user + 10K/min global ceiling, configurable per-skill
Input Firewall ABSENT in 29% -- Raw input to tools 5-layer Airgap prompt firewall with canary token detection
Output Sanitization ABSENT in 61% -- Raw output from tools Response sanitization + prompt injection detection
Transport Security PARTIAL -- Most support HTTPS Mandatory TLS, origin validation, no HTTP fallback
Memory Isolation ABSENT -- Shared process memory Per-user isolation with provenance + NIST 800-88 secure deletion
Supply Chain 47% missing lockfiles Rust binary, cargo-audit, Clippy, MSRV 1.91, 0 unsafe blocks
CMMC Adherence 0 servers claim adherence 17/17 CMMC Level 1 practices MET with documented evidence
Red Team Testing 0 servers publish red team results 25 red team tests + 10 PQC tests + 373 total tests + 10 CI gates
Post-Quantum Cryptography 0 servers -- Ed25519 only, vulnerable to Shor's algorithm Hybrid Ed25519 + ML-DSA-65 (FIPS 204) -- NIST Level 3 quantum resistance. Dual-family hashing (SHA-256 + SHA3-256). Both algorithms must verify independently.
Crypto Agility 0 servers -- Hardcoded algorithms, no upgrade path Three security levels (Classical, Hybrid3, Hybrid5). Algorithm selection per-signature. Backward-compatible with pre-PQC signatures. ML-DSA-87 (NIST Level 5) for government/military.

ddot Certification Program

The ddot Agent Security Audit provides instant, free security scoring for any MCP server or AI agent. For production deployments, ddot Certification verifies that agents operate through the ddot security gateway with full Wasm sandbox isolation, cryptographic signing, and tamper-evident audit trails.

ddot.build | The SSL of Agentic AI

Quantum Readiness: The Next Horizon

Every MCP server in this assessment uses classical cryptography vulnerable to quantum computers. ddot is the first agentic AI security protocol to implement post-quantum defenses.

The Threat

Shor's algorithm, running on a sufficiently large quantum computer, breaks Ed25519 in polynomial time. Every signature, every key exchange, every identity verification in the MCP ecosystem becomes forgeable.

This is not theoretical. NIST finalized post-quantum standards (FIPS 203/204/205) in 2024 precisely because the threat timeline is measured in years, not decades. Data harvested today under "harvest now, decrypt later" attacks will be exposed when quantum capability arrives.

0 of 1,693 MCP servers implement any form of post-quantum cryptography.

ddot's Defense

ddot implements hybrid classical + post-quantum signing using the belt-and-suspenders approach recommended by NIST SP 800-227:

Classical LayerEd25519 (FIPS 186-5)
Quantum LayerML-DSA-65 (FIPS 204, lattice-based)
Hash FamiliesSHA-256 + SHA3-256 (dual-family)
VerificationBoth algorithms must pass independently

If quantum computers break Ed25519, ML-DSA holds. If a cryptanalytic breakthrough weakens ML-DSA, Ed25519 holds. An attacker must break both algorithms simultaneously.

Security Levels

Level Algorithms NIST Category Signature Size Use Case
Classical Ed25519 only Level 1 64 B Legacy backward compatibility only
Hybrid3 Ed25519 + ML-DSA-65 Level 3 3,373 B Default for all new signatures
Hybrid5 Ed25519 + ML-DSA-87 Level 5 4,691 B Government, defense, critical infrastructure
Proven // 10 PQC Tests // All Passing // cargo test -p ddot-pqc
PQC-01 Hybrid L3 sign+verifyPASS
PQC-02 Hybrid L5 sign+verifyPASS
PQC-03 Classical backward compatPASS
PQC-04 Tampered binary detectionPASS
PQC-05 Min security level gatePASS
PQC-06 Expired key rejectionPASS
PQC-07 CRL key revocationPASS
PQC-08 Dual-family hash independencePASS
PQC-09 Serialization round-tripPASS
PQC-10 FIPS 204 size conformancePASS
test result: ok. 10 passed; 0 failed; 0 ignored -- finished in 0.03s

"The rest of the industry is building security for 2026. ddot is building security for 2036. When quantum computing renders every Ed25519 signature in the MCP ecosystem worthless, ddot-signed skills will still verify."

-- ddot Labs, on why post-quantum cryptography is non-negotiable

Recommendations

Based on the analysis of 1,693 MCP servers, ddot Labs recommends the following immediate actions for the ecosystem.

P0
Mandatory Execution Sandboxing
Every MCP server must sandbox tool execution. Wasm (Wasmtime/WasmEdge) provides the strongest model. The current state -- 64% of servers running with full OS privileges -- is untenable.
P0
Authentication by Default
The MCP specification should mandate transport-level authentication. 53% of servers accept anonymous connections. MCP clients must verify server identity before invoking tools.
P1
Ecosystem-Wide Audit Trail Standard
85% of servers have no audit logging. The MCP protocol should define a standard audit event format so that every tool invocation is logged with caller identity, parameters, and results.
P1
SDK Security Primitives
The official MCP SDKs must ship with built-in auth, rate limiting, input validation, and audit hooks. Server developers should inherit security by default, not build it from scratch.
P2
Industry Security Certification
Enterprise buyers need a way to verify that the MCP servers they deploy meet minimum security standards. Third-party certification programs must emerge to fill this gap.
P2
Responsible Disclosure Coordination
121 servers contain hardcoded secrets visible on public GitHub. Coordinated disclosure and credential rotation should be prioritized for the most critical exposures.
ddot