MCP Security Architecture Guide
Read the foundational guide on hardening Anthropicโs Model Context Protocol against tool poisoning and rogue servers. Explore MCP Security Guide โ
Hermes Codex independently benchmarks every mainstream MCP server across four critical dimensions:
Explore the interactive registry below to audit your agentic dependencies, filter by category or blast radius tier, and export your stack configuration.
Direct filesystem access allowing agents to read and mutate directory structures. Vulnerable to symlink escapes, path traversal, and sensitive config harvesting.
Enforce chroot or Linux container mount namespaces. Run with read-only root mount and explicit whitelist of disposable subdirectories.
Arbitrary shell execution on host. Direct RCE blast radius with access to ambient environment variables and process memory.
Mandatory ephemeral MicroVM (Firecracker/gVisor). Disallow host subshell access completely; strip parent env secrets.
Interacts with Docker daemon socket. Unsanitized volume mounts enable immediate root breakout onto the Docker host.
Rootless Docker or Podman only. Prohibit mounting host / or docker.sock inside spawned containers.
Cluster-wide orchestration via kubeconfig. Agent prompt injection can trigger unauthorized pod deployments or secret dumping.
RBAC scoping to dedicated non-privileged namespace. Disable secret viewing and pod exec verbs.
Allocates interactive PTY sessions. Threat actors use blind command injection to spawn reverse shells.
Attach strict seccomp filters and drop CAP_SYS_ADMIN. Restrict stdout/stderr buffer lengths.
Executes SQL statements on PostgreSQL databases. Susceptible to indirect prompt injection driving second-order SQL injection.
Read-only connection credentials. Block DDL commands (DROP, ALTER, CREATE) and use parameterized statement wrappers.
Local SQLite database operations. ATTACH DATABASE commands allow writing outside intended storage files.
Open in SQLITE_OPEN_READONLY mode unless explicit write transaction required. Disable ATTACH DATABASE.
MySQL queries and table scans. Threat actors inject SELECT INTO OUTFILE to plant web shells.
Disable FILE privilege. Require human approval for any non-SELECT statements.
Redis in-memory caching and session store. FlushAll or Config Set commands risk denial-of-service and host file overwrite.
Rename or disable dangerous commands (CONFIG, FLUSHALL, FLUSHDB, EVAL). Namespace agent keys with prefixes.
MongoDB NoSQL queries. NoSQL operator injection ($where, $regex) allows bypassing filters and exfiltrating collections.
Disallow $where JavaScript evaluation on MongoDB server. Enforce read-only collections.
Enterprise data warehouse querying. Prompt injection can trigger mass table unloads to untrusted S3 buckets.
Block COPY INTO @external_stage commands. Restrict warehouse credits and query duration timeouts.
Executes BigQuery analytics. Rogue queries can incur denial-of-wallet costs through petabyte-scale table scans.
Enforce maximum bytes billed quotas per query. Assign BigQuery Data Viewer role without job creation permissions.
Cypher queries on knowledge graphs. APOC procedure calls can allow host filesystem reading or network requests.
Disable apoc.import and apoc.export procedures. Run with read-only Cypher endpoint.
GitHub API automation. Allows commits, branch pushes, and CI workflow triggering. Attacker can poison source code or backdoors into master branch.
Strictly prohibit push to default branch. Enforce Branch Protection Rules requiring 2 human approvals for agent-authored PRs.
GitLab repository & CI pipeline control. Risk of harvesting CI/CD masked variables and secret tokens.
Scope Personal Access Token without admin:api. Prevent reading masked project CI variables.
Local git repository inspections. Submodule configurations and .git/hooks can execute arbitrary shell scripts on checkout.
Disable git hooks via git -c core.hooksPath=/dev/null. Run in isolated git worktree.
Sentry issue retrieval. Stack traces often contain embedded API secrets, JWT tokens, and user PII.
Scrub server-side sensitive variables in Sentry settings before passing issues to agent context.
Datadog observability queries. Log searches can expose plain-text passwords or internal infrastructure topology.
Filter sensitive log indexes. Limit API key permissions to metrics and dashboard reading.
Web search engine query execution. Primary conduit for indirect prompt injection via untrusted search snippets.
Sanitize returned snippets through dual-channel LLM verifier before blending into agent system memory.
Generic HTTP GET/POST fetch tool. High SSRF risk targeting cloud metadata services (169.254.169.254) and internal networks.
Block RFC 1918 private IP ranges and cloud metadata endpoints (169.254.169.254) at network and DNS levels.
Headless Chromium browser automation. Malicious web pages can trigger zero-day sandbox escapes or steal session cookies.
Launch Chrome with --no-sandbox forbidden in production; run within disposable MicroVM with isolated network interface.
Multi-browser web testing automation. Similar to Puppeteer, browser contexts can be subverted via untrusted JavaScript payloads.
Disable file downloads. Restrict navigation to approved domains via DNS proxy filtering.
AI-optimized web search. Susceptible to search result adversarial poisoning targeting agent decision trees.
Validate parsed search entities against internal ground truth before performing operational actions.
Slack bot interaction. Adversaries posting messages in monitored channels can inject indirect prompt instructions to hijack agent.
Restrict agent bot to dedicated testing channel. Never allow posting with user-impersonation tokens.
Google Drive file search and download. Poisoned shared Google Docs can inject persistent backdoor instructions.
Strictly isolate corporate drives by department. Use OAuth service accounts with read-only view scopes.
Notion workspace page and database sync. Injection inside Notion blocks can trigger unauthorized exfiltration.
Scope Notion integration token to designated sub-pages only. Prohibit editing workspace settings.
Linear task tracking. Malicious issue descriptions can manipulate agent prioritizations and sprint planning.
Limit API token to read-only on sensitive projects. Disallow deleting issues.
Jira and Confluence corporate knowledge integration. Widespread enterprise blast radius across internal tickets and documentation.
Enforce principle of least privilege on Atlassian API tokens. Restrict access to internal security boards.
Discord community bot operations. Untrusted user DMs can hijack customer-facing agents and broadcast phishing links.
Sanitize Discord user input with strict content moderation before agent processing. Block sending external URLs.
IMAP mailbox ingestion. Inbound phishing emails with hidden invisible text prompt injections can seize full control of the agent.
Strip HTML comments, invisible zero-width unicode, and CSS display:none text prior to LLM reasoning.
Outbound SMTP email dispatch. Hijacked agents can be utilized as spam relays or executive BEC impersonation channels.
Require out-of-band cryptographic approval for any email dispatched to external recipient domains.
Persistent knowledge graph memory. Poisoned memories persist across sessions, establishing durable agent backdoors.
Implement periodic memory auditing and provenance signatures on every stored entity and relation.
Vector database operations. Attacker injects adversarial embeddings to hijack retrieval-augmented generation (RAG).
Namespace vector spaces by tenant and verify cosine anomaly thresholds on newly upserted vectors.
Qdrant vector search. Payload tampering can manipulate filter logic and retrieve unauthorized records.
Sign and validate payload metadata with tamper-evident HMAC keys.
Weaviate vector store with GraphQL capabilities. Schema injection allows modifying class definitions.
Disable schema mutation APIs during production runtime.
Chroma embedding database. Local SQLite-backed deployments can suffer file locking or local database corruption.
Enforce client-side cryptographic hashing on text documents before embedding.
Broad AWS management capability. Hijacked agents can deploy crypto-miners or delete corporate S3 backups.
Enforce tight AWS IAM Service Control Policies (SCPs) and require MFA token generation for mutating actions.
Google Cloud Platform management. Risk of IAM privilege escalation via service account token creation.
Restrict iam.serviceAccountTokenCreator permissions. Use short-lived OAuth access tokens with 5-minute TTL.
Azure Resource Manager & Entra ID control. Agent compromise exposes cloud tenant directory and key vault secrets.
Lock down Key Vault access policies. Enforce Azure Privileged Identity Management (PIM) with approval workflows.
Cloudflare edge management. Attacker can subvert DNS records or disable WAF rules to expose origin servers.
Separate DNS administration token from CDN caching token. Restrict zone DNS modification rights.
Salesforce enterprise CRM operations. Direct exposure to confidential sales pipelines, customer contracts, and lead records.
Restrict SOQL queries to active operational fields. Mask all customer credit card and banking fields.
HubSpot marketing and contact automation. Risk of unauthorized mass marketing dispatch or PII database dump.
Scope private app token to read-only contacts. Disallow email marketing batch API dispatch.
Zendesk customer support ticketing. Attacker can read internal escalations or post malicious phishing replies.
Enforce human review on public replies. Redact customer passwords and auth tokens automatically.
Stripe payments and financial transactions. Severe kinetic blast radius: rogue refund issuances or unauthorized payments.
Use restricted API keys with read-only charges. Require physical hardware token confirmation for any transfer.
Obot enterprise tool catalog. Directly impacted by CVE-2026-105138 allowing regular users to dump static tool secrets.
Upgrade to Obot >= 0.26.2 immediately. Enforce RBAC on MCP catalog endpoints and migrate to ephemeral tokens.
Local model serving and execution. Model pulling can download poisoned GGUF weights with embedded exploits.
Verify SHA256 checksums of model weights against verified vendor manifests before loading into memory.
Hugging Face Hub interaction. Loading untrusted pickle-based models causes arbitrary Python code execution.
Enforce safetensors format only; block loading PyTorch pickle files completely.
Telegram messaging bot operations. Subject to IPC record injection and malicious attachment harvesting.
Sanitize incoming chat messages and block automatic attachment download without manual inspection.
Tmux session manager. Injects keystrokes directly into active terminal panes, bypassing shell argument filters.
Prohibit attaching to existing human user sessions. Run within dedicated ephemeral pty container.
Obsidian markdown note vault access. Threat actors inject hidden markdown links and exfiltration payloads into daily notes.
Confine vault access to a sandboxed read-only clone. Strip external URI schemes from generated notes.
MCP Security Architecture Guide
Read the foundational guide on hardening Anthropicโs Model Context Protocol against tool poisoning and rogue servers. Explore MCP Security Guide โ
Declare Your Security Stack (MyStack)
Declare your deployed AI frameworks, inference backends, and MCP tools to evaluate real-time exposure. Open MyStack Declaration โ
AI Agent Blast Radius Simulator
Simulate what an attacker can access when an agent identity or MCP connector is compromised. Run Blast Radius Simulator โ
Agentic Attack Patterns
Analyze verified adversarial patterns including function hijacking, semantic rug-pulls, and agent poisoning. Browse Attack Patterns โ