What mcp-audit checks
mcp-audit can report 225 finding IDs, plus 6 families of IDs generated at scan time. They come from 8 built-in analyzers, 40 community rules that run by default and 91 Semgrep rules for MCP server source code.
Severities come from the severity framework and the rule files; OWASP codes from the mapping file.
Repository files that run code 8 checks
Files a cloned repository can use to start programs when you open or trust it: project MCP configs, VS Code tasks that run when the folder opens, auto-run files for Gemini CLI, Codex and Claude Code, git settings and hooks that run a program when an agent runs git, and symlinks that point somewhere else.
Runs with mcp-audit scan --project .
| ID | What it flags | Severity | OWASP MCP |
|---|---|---|---|
TRUST-001 | MCP server defined in project-level config (TrustFall risk) | HIGH | MCP09 |
TRUST-002 | Project-scoped config path is a symlink | HIGH, MEDIUM or INFO | MCP09 MCP05 |
TRUST-003 | Repo-planted IDE auto-execution file | CRITICAL or HIGH | MCP05 MCP09 |
TRUST-004 | Config path is a symlink (user-global / explicit) | INFO | MCP09 |
TRUST-005 | Symlinked directory was not scanned | LOW | MCP09 |
TRUST-006 | Git config runs a command | HIGH, MEDIUM, LOW or INFO | MCP05 |
TRUST-007 | Git repository shape runs repo-controlled code | HIGH or INFO | MCP05 |
TRUST-008 | Repo-planted auto-run file of another coding-agent harness | CRITICAL, HIGH or MEDIUM | MCP05 MCP09 |
AI agents in CI workflows 3 checks
GitHub Actions workflows where an AI coding agent step reads text a stranger can write, such as an issue or a comment, while it holds secrets, write access and tools that can run commands. Workflow files are parsed as data and never run.
Runs with mcp-audit check mcp-audit scan --project .
| ID | What it flags | Severity | OWASP MCP |
|---|---|---|---|
WORKFLOW-001 | CI agent runs on text a stranger can write, with secrets, write access and unrestricted tools | HIGH | MCP06 MCP01 |
WORKFLOW-002 | CI agent's only tool restriction is a blocklist | MEDIUM | MCP06 MCP01 |
WORKFLOW-003 | Gemini CLI step exposed to the CVSS 10.0 workspace-trust flaw (GHSA-wpqr-6v78-jr5g) | HIGH | MCP05 MCP04 |
Coding-agent settings and permission rules 39 checks
Claude Code settings that skip permission prompts, approve every project server, or pre-approve commands that can write or reach the network. Permission rules that Claude Code ignores or that are not a boundary. Sandbox settings a repository can loosen. Hook commands that reach the network or touch agent config files. Servers that your organisation's managed Claude Code policy would block, policy set only where the user can change it, safety settings that got weaker since the last scan, editor extension allowlists that are not enforced, and credential files a session can read without a prompt.
Runs with mcp-audit check mcp-audit scan --project .
| ID | What it flags | Severity | OWASP MCP |
|---|---|---|---|
CFHYG-005 | Claude Code hooks section detected in config | MEDIUM | MCP01 MCP07 |
EXTPOL-001 | Editor extension allowlist inventory | INFO | MCP04 |
EXTPOL-002 | Editor extension allowlist is set only in user settings | LOW | MCP04 |
EXTPOL-003 | Editor extension allowlist policy allows everything | LOW | MCP04 |
GUARD-001 | Agent guardrail loosened since the last scan | HIGH or MEDIUM | MCP02 |
GUARD-002 | Managed settings file changed since the last scan | INFO | MCP02 |
HARNESS-001 | Claude Code sessions start with permission prompts skipped | HIGH or MEDIUM | MCP02 MCP07 |
HARNESS-002 | Every project .mcp.json server is approved without a prompt | HIGH | MCP09 MCP07 |
HARNESS-003 | Confirmation before bypass mode is switched off | MEDIUM | MCP02 MCP07 |
HARNESS-004 | Every shell command is pre-approved | HIGH | MCP02 MCP05 |
HARNESS-005 | Allow rules pre-approve commands that can write or reach the network | MEDIUM | MCP02 MCP05 |
HARNESS-006 | Approval is off and nothing guards recursive deletes | MEDIUM | MCP02 MCP05 |
HARNESS-007 | Named .mcp.json servers are approved without a prompt | INFO | MCP09 MCP07 |
HOOK-001 | Hook command contains network-egress instruction | HIGH | MCP05 MCP07 |
HOOK-002 | Hook command references agent config file path | HIGH | MCP05 MCP07 |
HOOK-004 | Guard hook command may never run on Windows | MEDIUM or LOW | MCP02 MCP07 |
HOOK-005 | Cursor hooks file can't be read, so its hooks never run | MEDIUM | MCP02 MCP07 |
MANAGED-001 | Managed MCP policy on this machine | INFO | MCP09 |
MANAGED-002 | Configured MCP server that the managed policy keeps from loading | INFO | MCP09 |
MANAGED-003 | Managed settings are deployed but set no MCP policy | LOW | MCP09 |
MANAGED-004 | Settings file adds servers to an unlocked managed MCP allowlist | MEDIUM or LOW | MCP02 MCP09 |
MANAGED-005 | Managed MCP allowlist admits servers by name only | LOW | MCP09 MCP07 |
MANAGED-006 | Claude Code policy is set, not enforced | MEDIUM | MCP09 MCP07 |
MANAGED-007 | Claude Code admin policy source is not a valid settings object | INFO | MCP09 |
PERM-001 | Permission rules that Claude Code ignores or never matches | HIGH, MEDIUM or INFO | MCP07 MCP02 |
PERM-002 | Path rules with one leading slash point inside the project, not at / | MEDIUM or INFO | MCP07 MCP10 |
PERM-003 | Allow rules whose trailing * also matches other programs | LOW | MCP02 MCP05 |
PERM-004 | Deny or ask rules on network or delete commands are not a boundary without the sandbox | MEDIUM or INFO | MCP05 MCP07 |
PERM-005 | WebFetch(domain:*) allow opens the sandbox to every host | HIGH or LOW | MCP02 MCP10 |
PERM-006 | A bare Bash ask rule is skipped for sandboxed commands | MEDIUM | MCP07 MCP05 |
PERM-007 | Allow rule pre-approves every tool of a server with high-severity findings | HIGH | MCP02 MCP07 |
REACH-001 | Credential files a Claude Code session can read without asking | LOW | MCP01 |
REACH-002 | Claude Code's login store is readable by other local accounts | MEDIUM | MCP01 |
REACH-003 | .env files in the working directory with no Read deny rule | LOW | MCP01 |
SANDBOX-001 | Repository settings weaken or switch off the Bash sandbox | HIGH, MEDIUM or INFO | MCP02 MCP05 |
SANDBOX-002 | Repository settings take code-running commands out of the sandbox | HIGH or MEDIUM | MCP05 MCP02 |
SANDBOX-003 | Repository settings let sandboxed commands reach hosts anyone can publish to | MEDIUM or INFO | MCP02 MCP10 |
SANDBOX-004 | Sandbox keys this settings file can't set (ignored by Claude Code) | INFO | MCP02 |
SANDBOX-005 | Local Claude Code settings are not ignored by git | MEDIUM or LOW | MCP10 MCP02 |
Agent instruction files, skills and subagents 22 checks
Text the agent reads as instructions: Claude Code commands, CLAUDE.md and AGENTS.md files, the files they import, Cursor rules, and Copilot instruction and prompt files. mcp-audit looks for injected orders, hidden characters, outside links and heartbeat files that hide a step. It reads every file in a Claude Code skill, not only SKILL.md, for install steps that run remote code. It reads subagent definitions for looser permissions, hooks and their own MCP servers, and reports when an instruction file changed since the last scan.
Runs with mcp-audit scan --project . --include-agent-files mcp-audit agent-files scan
| ID | What it flags | Severity | OWASP MCP |
|---|---|---|---|
AGENT-001 | Subagent asks for a looser permission mode | MEDIUM or LOW | MCP02 MCP07 |
AGENT-002 | disallowedTools specifier removes the whole tool | INFO | MCP02 |
AGENT-003 | Subagent defines hooks that run shell commands | MEDIUM or INFO | MCP05 MCP07 |
AGENT-004 | Subagent declares its own MCP server | MEDIUM, LOW or INFO | MCP09 |
INSTR-001 | Instruction file changed since the last scan | INFO | MCP10 |
INSTR-002 | Instruction file changed and the change brought a finding | MEDIUM | MCP10 MCP03 |
MEM-001 | Imperative override instruction in memory file | MEDIUM | MCP03 |
MEM-002 | Injection pattern in memory file | MEDIUM | MCP03 |
MEM-003 | Instruction file imports a file from outside the project / a .gitignore'd file | MEDIUM or LOW | MCP10 |
MEM-004 | Heartbeat file hides what the agent does | HIGH | MCP06 MCP08 |
MEM-005 | Instruction file holds a conditional sleeper trigger | MEDIUM | MCP03 MCP06 |
MEM-006 | Instruction file tells the agent to send secrets out | HIGH or MEDIUM | MCP03 MCP01 |
MEM-007 | Agent instruction file in Downloads or Desktop | MEDIUM | MCP03 MCP06 |
MEM-008 | Instruction file changes machine-wide agent settings or approvals | HIGH or MEDIUM | MCP03 MCP02 |
SKILL-001 | Injection/exfiltration instruction in skill file | HIGH or INFO | MCP03 MCP01 |
SKILL-002 | Obfuscated or oversized block in skill/command file | MEDIUM | MCP03 |
SKILL-003 | Skill file references external URL | MEDIUM | MCP03 MCP04 |
SKILL-004 | Skill bundles executable scripts | INFO | MCP03 |
SKILL-005 | Skill contains an install step that runs remote or hidden code | HIGH, LOW or INFO | MCP04 MCP05 |
SKILL-006 | Skill ships something Claude Code runs, or tells the agent to drop its guardrails | HIGH, MEDIUM or LOW | MCP01 MCP02 MCP04 MCP05 MCP07 MCP09 |
SKILL-007 | Skill contains content that cannot be inspected | MEDIUM or LOW | MCP03 MCP04 |
SKILL-008 | Skill bundles code and has another finding | HIGH | MCP04 MCP05 |
Claude Code plugins and mods 10 checks
Installed plugins whose recorded version or checkout does not match their marketplace pin, or that follow an unpinned branch, and mods that can approve tool calls, read secrets and send data out, change what Claude reads, or start programs. Mods are read, never run.
Runs with mcp-audit check mcp-audit plugins scan
| ID | What it flags | Severity | OWASP MCP |
|---|---|---|---|
MOD-001 | Mod can approve tool calls before the permission prompt | HIGH | MCP07 MCP04 |
MOD-002 | Mod can read secrets or files and send data out | HIGH or MEDIUM | MCP01 MCP04 |
MOD-003 | Mod can change what Claude reads | MEDIUM | MCP03 MCP06 |
MOD-004 | Mod starts programs or changes their environment | MEDIUM | MCP05 MCP04 |
MOD-005 | Claude Code mod installed | INFO | MCP04 |
MOD-006 | Plugin directories load in every Claude Code session | MEDIUM | MCP04 |
PLUGIN-001 | Installed plugin does not match its marketplace pin | MEDIUM | MCP04 |
PLUGIN-002 | Plugin source is not pinned to a commit | LOW | MCP04 |
PLUGIN-003 | Plugin checkout is not the commit its marketplace pins | CRITICAL or MEDIUM | MCP04 |
PLUGIN-004 | Plugin update channel cannot hold its pin | HIGH, MEDIUM or LOW | MCP04 |
MCP server configs 20 checks
How each configured server is launched and reached: secrets written into the config, plain-HTTP endpoints, sudo launches, unpinned package fetches, remote servers with no authentication, and config files other users can read or write. If a check crashes on a server, the scan reports it instead of passing that server as clean.
Runs with mcp-audit check
| ID | What it flags | Severity | OWASP MCP |
|---|---|---|---|
AUTH-001 | Remote MCP server configured without authentication | HIGH or MEDIUM | MCP06 |
AUTH-002 | OAuth-configured server missing audience/resource binding | MEDIUM | MCP06 |
CFHYG-001 | MCP config file is world-readable | HIGH | MCP01 |
CFHYG-002 | MCP config file resides in world-writable directory | HIGH | MCP01 MCP09 |
CFHYG-003 | Config file stores inline plaintext secret | HIGH | MCP01 |
CFHYG-004 | Config uses environment variable references for credentials | INFO | MCP01 |
CFHYG-006 | ANTHROPIC_BASE_URL overrides Anthropic API endpoint | MEDIUM | MCP01 MCP08 |
CFHYG-007 | Project settings env redirects API traffic | HIGH | MCP01 MCP08 |
CFHYG-008 | Project settings send API requests to an Azure Foundry resource | INFO | MCP01 MCP08 |
CRED-001 | API key or secret in environment variable value | HIGH | MCP01 |
CRED-002 | API key or secret in command arguments | HIGH | MCP01 |
CRED-003 | Literal secret in authentication header | HIGH or INFO | MCP01 |
CRED-004 | Secret saved in a permission rule | HIGH or MEDIUM | MCP01 |
CRED-005 | Secret in a permission rule in git history | HIGH | MCP01 |
CRED-006 | Claude Code settings file copied into a container image or npm package | HIGH or LOW | MCP01 |
SCAN-ERR | Analyzer crashed on a server | HIGH | none |
TRANSPORT-001 | Unencrypted remote endpoint (HTTP) | MEDIUM | MCP07 |
TRANSPORT-002 | Elevated privilege execution (sudo/root) | HIGH | MCP02 MCP05 |
TRANSPORT-003 | Runtime package fetching without version pin | MEDIUM or LOW | MCP04 |
TRANSPORT-004 | Server binds to wildcard network interface (0.0.0.0) | HIGH | MCP07 |
Tool descriptions 29 checks
Orders hidden in the descriptions an agent reads: exfiltration instructions, prompt-format injection, behaviour overrides, invisible Unicode and lookalike characters. With --connect it reads live tool lists, flags a tool name claimed by two servers, and flags descriptions that name another server's tool or tell the agent to rewrite other tools' arguments. It also flags a tool marked read-only whose own inputs run commands or write.
Runs with mcp-audit check mcp-audit scan --connect
| ID | What it flags | Severity | OWASP MCP |
|---|---|---|---|
ANNOT-001 | Tool annotation contradicts the tool's own input schema | LOW | MCP03 MCP06 |
COLLIDE-001 | Tool name collision: tool claimed by multiple servers | HIGH or MEDIUM | MCP01 MCP09 |
DELEG-001 | Tool description names another server's tool | MEDIUM | MCP03 MCP06 |
DELEG-002 | Tool description tells the agent to act when something else happens | MEDIUM | MCP03 MCP06 |
DELEG-003 | Tool description tells the agent to rewrite other tools' arguments | HIGH | MCP03 MCP06 |
POISON-001 | Tool description contains SSH/credentials exfiltration instructions | CRITICAL | MCP03 MCP01 |
POISON-002 | Tool description contains cloud credentials exfiltration instructions | CRITICAL | MCP03 MCP01 |
POISON-003 | Tool description contains .env file exfiltration instructions | CRITICAL | MCP03 MCP01 |
POISON-010 | Tool description contains XML instruction injection | HIGH | MCP03 MCP06 |
POISON-011 | Tool description contains LLM prompt format injection markers | HIGH | MCP03 MCP06 |
POISON-012 | Tool description contains behavioral override instruction | HIGH | MCP03 MCP06 |
POISON-020 | Tool description contains data exfiltration via encoding or side channel | HIGH | MCP03 MCP10 |
POISON-021 | Tool description contains hidden parameter exfiltration | HIGH | MCP03 MCP10 |
POISON-030 | Tool description contains cross-tool manipulation instruction | MEDIUM | MCP03 MCP06 |
POISON-040 | Tool description contains zero-width Unicode characters | MEDIUM | MCP03 |
POISON-041 | Concealed instruction via concealment channel | HIGH | MCP03 MCP01 |
POISON-042 | Concealment channel present, no matched payload | LOW | MCP03 |
POISON-043 | Zero-width binary channel present, no matched payload | MEDIUM | MCP03 |
POISON-044 | Variation-selector channel present, no matched payload | MEDIUM | MCP03 |
POISON-045 | Text hidden below blank-line padding | HIGH or MEDIUM | MCP03 |
POISON-050 | Excessive tool description length | LOW | MCP03 |
POISON-060 | Tool description contains Unicode homoglyph characters | HIGH | MCP03 |
POISON-070 | Secrecy directive | HIGH | MCP03 MCP06 |
POISON-071 | Autonomy override directive | HIGH or LOW | MCP03 MCP06 |
POISON-072 | Ask-the-user tool disabled | MEDIUM | MCP03 MCP06 |
POISON-073 | Coercive framing | LOW | MCP03 MCP06 |
POISON-074 | Startup execution directive | LOW | MCP03 MCP05 |
POISON-075 | Remote instruction following | LOW | MCP03 MCP06 |
POISON-076 | Oversight bypass directive | HIGH | MCP03 MCP06 |
Dangerous combinations of servers 12 checks
Servers that are harmless alone but together let an injected prompt read data and send it out, including by posting to GitHub, Slack or Jira, or write a file another server can run. killchain ranks the changes that break the most paths.
Runs with mcp-audit check mcp-audit killchain
| ID | What it flags | Severity | OWASP MCP |
|---|---|---|---|
INTEG-001 | Dangerous capability combination: file write + shell execution (plant-then-execute) | HIGH | MCP05 |
TOXIC-001 | Dangerous capability combination: file read + network exfiltration | HIGH | MCP05 MCP10 |
TOXIC-002 | Dangerous capability combination: file read + email exfiltration | HIGH | MCP05 MCP10 |
TOXIC-003 | Dangerous capability combination: secrets store + network access | CRITICAL | MCP01 MCP10 |
TOXIC-004 | Dangerous capability combination: file read + shell execution | HIGH | MCP10 |
TOXIC-005 | Dangerous capability combination: database + network | HIGH | MCP05 MCP10 |
TOXIC-006 | Dangerous capability combination: shell execution + network | CRITICAL | MCP05 MCP10 |
TOXIC-007 | Capability combination: git access + network | MEDIUM | MCP10 |
TOXIC-008 | Dangerous capability combination: file read + comms posting | HIGH | MCP05 MCP10 |
TOXIC-009 | Dangerous capability combination: secrets store + comms posting | CRITICAL | MCP01 MCP10 |
TOXIC-010 | Dangerous capability combination: database + comms posting | HIGH | MCP05 MCP10 |
TOXIC-011 | Dangerous capability combination: shell execution + comms posting | CRITICAL | MCP05 MCP10 |
Packages and supply chain 19 checks
Lookalike package names, unscoped twins of known servers, packages listed as malicious in OSV, recorded advisories for a package, deprecated versions, and missing or mismatched Sigstore provenance. Provenance, hash and CVE checks use the network, so they are opt-in.
Runs with mcp-audit vet <package> mcp-audit scan --verify-signatures
| ID | What it flags | Severity | OWASP MCP |
|---|---|---|---|
ATTEST-010 | Sigstore provenance verified (valid match) | INFO | none |
ATTEST-011 | Sigstore attestation from unexpected publisher | HIGH | MCP04 |
ATTEST-012 | Invalid Sigstore signature (tampered or forged) | CRITICAL | MCP04 |
ATTEST-013 | Expected Sigstore attestation missing for known-provenance package | MEDIUM | MCP04 |
ATTEST-014 | No Sigstore attestation found | MEDIUM or INFO | MCP04 |
ATTEST-015 | Sigstore attestation verification error | INFO | none |
SC-001 | Package name is one edit distance from a known-good package (typosquat) | CRITICAL | MCP04 |
SC-002 | Package name is two edits from a known-good package | HIGH | MCP04 |
SC-003 | Package name is three edits from a known-good package | MEDIUM | MCP04 |
SC-004 | Known CVE advisory recorded for this package | HIGH or MEDIUM | MCP04 |
SC-005 | Deprecated package resolved at lock time | MEDIUM | MCP04 |
SC-007 | Unscoped twin of a known package | MEDIUM | MCP04 |
SC-008 | Package listed as malicious in OSV (MAL-* record) | CRITICAL or MEDIUM | MCP04 |
SC-009 | Known CVE fix not comparable with pinned version | INFO | MCP04 |
SC-010 | Archived MCP server, no security fixes coming | MEDIUM | MCP04 |
SC-011 | Known malicious package in the npx cache | CRITICAL | MCP04 |
VULN-UNPINNED | Unpinned version (scan --check-vulns) | LOW | MCP04 |
VULN-UNRESOLVED | Vulnerability lookup incomplete (scan --check-vulns) | INFO | MCP04 |
ATT-… | Hash attestation finding (Layer 1 --verify-hashes). (ID generated at scan time) | varies | MCP04 |
Approved servers and drift 24 checks
What changed since you approved a setup: servers added, removed or launched differently, a hand-edited lock file, and, with --connect, tools added or changed since approval. A server started through mcp-audit run is checked against the lock as it launches and while it runs, and a policy can require every server to be wrapped that way.
Runs with mcp-audit lock --verify mcp-audit check mcp-audit wrap
| ID | What it flags | Severity | OWASP MCP |
|---|---|---|---|
BL-001 | Baseline file could not be parsed | not rated | none |
LOCK-001 | Server drifted from lock | HIGH | MCP04 |
LOCK-002 | Server not in lock | HIGH | MCP04 |
LOCK-003 | Locked server missing | MEDIUM | MCP04 |
LOCK-004 | Resolution drifted | CRITICAL or HIGH | MCP04 |
LOCK-005 | Lock file tampered or hand-edited | CRITICAL | MCP04 |
LOCK-006 | Tool added since lock | HIGH | MCP02 MCP04 |
LOCK-007 | Tool contract changed since lock | HIGH | MCP03 MCP04 |
LOCK-008 | Tool description changed since lock | MEDIUM | MCP03 MCP04 |
LOCK-009 | Tool removed since lock | LOW | MCP04 |
LOCK-010 | Resolution context changed | INFO | MCP04 |
LOCK-011 | Lock section digest does not hold | HIGH | MCP04 |
LOCK-012 | Lock file is not in canonical form | LOW | MCP04 |
LOCK-013 | Launch arguments carry a credential | LOW | MCP04 |
LOCK-014 | Repository guardrail file loosened since the lock | HIGH | MCP02 |
RUGPULL-000 | First scan — baseline recorded for server | INFO | none |
RUGPULL-001 | Server configuration changed since last scan | HIGH | MCP03 MCP04 |
RUGPULL-002 | New MCP server detected since last scan | INFO | MCP09 |
RUGPULL-003 | Previously tracked server no longer configured | INFO | none |
RUN-001 | No lock covers this server | HIGH | MCP04 |
RUN-002 | Lock unreadable at launch | HIGH | MCP04 |
RUN-003 | No tool record for a launched server | HIGH | MCP04 |
WRAP-001 | Server not launched through mcp-audit run | HIGH | MCP04 |
DRIFT-… | Baseline drift finding. (ID generated at scan time) | varies | MCP09 |
Community rules 41 checks
YAML rules in an open format that run on every scan: risky launch commands, inline code, shell metacharacters, sensitive variable names, sampling, and more.
Run on every scan. mcp-audit rule list shows the rules that are loaded.
40 of 41 community rules run by default; STDIO-001 ships disabled. The rule format is Apache 2.0 and open to contributions.
| ID | What it flags | Severity | OWASP MCP |
|---|---|---|---|
COMM-001 | Prohibited network binary used as MCP server | HIGH | MCP05 |
COMM-002 | curl used as MCP server binary | MEDIUM | MCP05 |
COMM-003 | Browser sandbox disabled | MEDIUM | MCP05 MCP10 |
COMM-004 | Unrecognized stdio server binary | LOW | MCP05 MCP09 |
COMM-005 | /tmp referenced in server args | MEDIUM | MCP05 |
COMM-006 | Inline code execution via -c flag | HIGH | MCP05 |
COMM-007 | eval or exec passed in server args | HIGH | MCP05 |
COMM-008 | SSH agent socket exposed to MCP server | MEDIUM | MCP01 |
COMM-009 | AWS session token exposed to MCP server | HIGH | MCP01 |
COMM-010 | npx used without pinned version | LOW | MCP04 |
COMM-011 | Broad Deno permission flags in use | MEDIUM | MCP02 |
COMM-012 | Server binds to all interfaces (0.0.0.0) | HIGH | MCP07 |
COMM-013 | npx auto-confirm flag (silent execution) | HIGH | MCP04 MCP05 |
COMM-014 | MCP Sampling capability declared | LOW | MCP04 MCP06 |
COMM-015 | Shell metacharacters in server args | CRITICAL | MCP05 |
COMM-016 | Role-claim poisoning in server args | HIGH | MCP01 |
COMM-017 | Authority-claim poisoning — AI identity impersonation | CRITICAL | MCP01 |
COMM-018 | Output-format hijack in server args | MEDIUM | MCP01 |
COMM-019 | Side-channel data exfiltration instruction in server args | HIGH | MCP01 MCP02 |
COMM-020 | Confabulation trigger in server args | HIGH | MCP01 |
COMM-021 | Env key name suggests private or secret key | HIGH | MCP02 |
COMM-022 | Env key name suggests access token | MEDIUM | MCP02 |
COMM-023 | Env key name suggests password | HIGH | MCP02 |
COMM-024 | HTTP endpoint on sensitive or production port | MEDIUM | MCP06 |
COMM-025 | Localhost endpoint with explicit SSE/HTTP transport URL | LOW | MCP06 MCP05 |
COMM-026 | Unencrypted WebSocket URL in server args | MEDIUM | MCP06 |
COMM-027 | Generic or unnamed server key | LOW | MCP07 |
COMM-028 | Command is an absolute path | MEDIUM | MCP04 |
COMM-029 | Command references a temp or volatile directory | HIGH | MCP04 |
COMM-030 | Generic-named server with network transport URL | LOW | MCP09 MCP07 |
COMM-031 | Remote MCP server configured without authentication | LOW | MCP06 |
COMM-033 | Project-level MCP config file present in repository | LOW | MCP09 |
COMM-034 | Overprivileged MCP server credential (potential god key) | MEDIUM | MCP06 |
COMM-035 | TLS certificate checking disabled by environment variable | MEDIUM | MCP01 |
COMM-036 | Container launched with host takeover flags | HIGH | MCP05 |
COMM-037 | Container launched on the host network | MEDIUM | MCP05 |
COMM-038 | Package installed from an extra package index | MEDIUM | MCP04 |
COMM-039 | Default package index replaced or custom npm registry configured | LOW | MCP04 |
STDIO-001 | Launcher binary not in allowlist (ships disabled) | MEDIUM | MCP04 MCP09 |
STDIO-002a | Hidden (dotfile) launcher binary | HIGH | MCP04 MCP05 |
STDIO-002b | Inline script launcher | HIGH | MCP05 |
More surfaces 4 checks
IDs generated at scan time for IDE extensions, MCP server source code, hash pins, baselines and your own governance policy.
Runs with mcp-audit extensions scan mcp-audit sast <path>
| ID | What it flags | Severity | OWASP MCP |
|---|---|---|---|
EXT-… | IDE extension security finding. (ID generated at scan time) | varies | MCP04 |
SAST-… | SAST finding from Semgrep rule execution. (ID generated at scan time) | varies | MCP01 MCP05 MCP07 |
GOV-… | Governance policy finding. (ID generated at scan time) | varies | none |
VULN-… | Known-vulnerability finding from scan --check-vulns (network, opt-in). (ID generated at scan time) | varies | MCP04 |
OWASP MCP Top 10
Findings carry their OWASP MCP Top 10 codes in terminal, JSON and SARIF output. The mapping is machine-readable (OWASP 2025-beta).
MCP01 Token Mismanagement and Secret Exposure 38 checks
CFHYG-001, CFHYG-002, CFHYG-003, CFHYG-004, CFHYG-005, CFHYG-006, CFHYG-007, CFHYG-008, COLLIDE-001, COMM-008, COMM-009, COMM-016, COMM-017, COMM-018, COMM-019, COMM-020, COMM-035, CRED-001, CRED-002, CRED-003, CRED-004, CRED-005, CRED-006, MEM-006, MOD-002, POISON-001, POISON-002, POISON-003, POISON-041, REACH-001, REACH-002, REACH-003, SKILL-001, SKILL-006, TOXIC-003, TOXIC-009, WORKFLOW-001, WORKFLOW-002
MCP02 Privilege Escalation via Scope Creep 31 checks
AGENT-001, AGENT-002, COMM-011, COMM-019, COMM-021, COMM-022, COMM-023, GUARD-001, GUARD-002, HARNESS-001, HARNESS-003, HARNESS-004, HARNESS-005, HARNESS-006, HOOK-004, HOOK-005, LOCK-006, LOCK-014, MANAGED-004, MEM-008, PERM-001, PERM-003, PERM-005, PERM-007, SANDBOX-001, SANDBOX-002, SANDBOX-003, SANDBOX-004, SANDBOX-005, SKILL-006, TRANSPORT-002
MCP03 Tool Poisoning 44 checks
ANNOT-001, DELEG-001, DELEG-002, DELEG-003, INSTR-002, LOCK-007, LOCK-008, MEM-001, MEM-002, MEM-005, MEM-006, MEM-007, MEM-008, MOD-003, POISON-001, POISON-002, POISON-003, POISON-010, POISON-011, POISON-012, POISON-020, POISON-021, POISON-030, POISON-040, POISON-041, POISON-042, POISON-043, POISON-044, POISON-045, POISON-050, POISON-060, POISON-070, POISON-071, POISON-072, POISON-073, POISON-074, POISON-075, POISON-076, RUGPULL-001, SKILL-001, SKILL-002, SKILL-003, SKILL-004, SKILL-007
MCP04 Software Supply Chain Attacks 62 checks
ATTEST-011, ATTEST-012, ATTEST-013, ATTEST-014, COMM-010, COMM-013, COMM-014, COMM-028, COMM-029, COMM-038, COMM-039, EXTPOL-001, EXTPOL-002, EXTPOL-003, LOCK-001, LOCK-002, LOCK-003, LOCK-004, LOCK-005, LOCK-006, LOCK-007, LOCK-008, LOCK-009, LOCK-010, LOCK-011, LOCK-012, LOCK-013, MOD-001, MOD-002, MOD-004, MOD-005, MOD-006, PLUGIN-001, PLUGIN-002, PLUGIN-003, PLUGIN-004, RUGPULL-001, RUN-001, RUN-002, RUN-003, SC-001, SC-002, SC-003, SC-004, SC-005, SC-007, SC-008, SC-009, SC-010, SC-011, SKILL-003, SKILL-005, SKILL-006, SKILL-007, SKILL-008, STDIO-001, STDIO-002a, TRANSPORT-003, VULN-UNPINNED, VULN-UNRESOLVED, WORKFLOW-003, WRAP-001
MCP05 Command Injection and Execution 45 checks
AGENT-003, COMM-001, COMM-002, COMM-003, COMM-004, COMM-005, COMM-006, COMM-007, COMM-013, COMM-015, COMM-025, COMM-036, COMM-037, HARNESS-004, HARNESS-005, HARNESS-006, HOOK-001, HOOK-002, INTEG-001, MOD-004, PERM-003, PERM-004, PERM-006, POISON-074, SANDBOX-001, SANDBOX-002, SKILL-005, SKILL-006, SKILL-008, STDIO-002a, STDIO-002b, TOXIC-001, TOXIC-002, TOXIC-005, TOXIC-006, TOXIC-008, TOXIC-010, TOXIC-011, TRANSPORT-002, TRUST-002, TRUST-003, TRUST-006, TRUST-007, TRUST-008, WORKFLOW-003
MCP06 Intent Flow Subversion 28 checks
ANNOT-001, AUTH-001, AUTH-002, COMM-014, COMM-024, COMM-025, COMM-026, COMM-031, COMM-034, DELEG-001, DELEG-002, DELEG-003, MEM-004, MEM-005, MEM-007, MOD-003, POISON-010, POISON-011, POISON-012, POISON-030, POISON-070, POISON-071, POISON-072, POISON-073, POISON-075, POISON-076, WORKFLOW-001, WORKFLOW-002
MCP07 Insufficient Authentication and Authorization 25 checks
AGENT-001, AGENT-003, CFHYG-005, COMM-012, COMM-027, COMM-030, HARNESS-001, HARNESS-002, HARNESS-003, HARNESS-007, HOOK-001, HOOK-002, HOOK-004, HOOK-005, MANAGED-005, MANAGED-006, MOD-001, PERM-001, PERM-002, PERM-004, PERM-006, PERM-007, SKILL-006, TRANSPORT-001, TRANSPORT-004
MCP09 Shadow MCP Servers 24 checks
AGENT-004, CFHYG-002, COLLIDE-001, COMM-004, COMM-030, COMM-033, HARNESS-002, HARNESS-007, MANAGED-001, MANAGED-002, MANAGED-003, MANAGED-004, MANAGED-005, MANAGED-006, MANAGED-007, RUGPULL-002, SKILL-006, STDIO-001, TRUST-001, TRUST-002, TRUST-003, TRUST-004, TRUST-005, TRUST-008
Supported clients 12 clients
mcp-audit finds these MCP config files on its own. Pass --path to scan any other file or folder.
| Client | Config file |
|---|---|
| Claude Desktop | ~/Library/Application Support/Claude/claude_desktop_config.json |
| Cursor | ~/.cursor/mcp.json |
| VS Code | .vscode/mcp.json (workspace) |
| Windsurf | ~/.codeium/windsurf/mcp_config.json |
| Claude Code (user) | ~/.claude.json |
| Claude Code (project) | .mcp.json (project root) |
| GitHub Copilot CLI | ~/.copilot/mcp-config.json |
| Augment Code | ~/.augment/settings.json |
| Amazon Q Developer (project) | .amazonq/mcp.json (project root; scan --project only) |
| Codex | $CODEX_HOME/config.toml (default ~/.codex/config.toml); .codex/config.toml (project) — TOML [mcp_servers.<name>] |
| Gemini CLI | ~/.gemini/settings.json; .gemini/settings.json (project) — mcpServers, comments allowed |
| OpenCode | ~/.config/opencode/opencode.json(c), $OPENCODE_CONFIG; opencode.json(c) (project) — mcp, JSONC |
Files read in a repository 42 paths
mcp-audit scan --project <dir> walks the tree for these paths. The injection and obfuscation checks on instruction files need --include-agent-files; imports, heartbeat files, skill packages, subagents, git settings and workflows are checked in a default scan. Symlinked directories are reported, never followed.
| Path | What it is | Client |
|---|---|---|
.mcp.json | MCP config or agent settings (servers start from it) | claude-code |
.claude/settings.json | MCP config or agent settings (servers start from it) | claude-code |
.claude/settings.local.json | MCP config or agent settings (servers start from it) | claude-code |
.cursor/mcp.json | MCP config or agent settings (servers start from it) | cursor |
.cursor/settings.json | MCP config or agent settings (servers start from it) | cursor |
.vscode/mcp.json | MCP config or agent settings (servers start from it) | vscode |
.amazonq/mcp.json | MCP config or agent settings (servers start from it) | amazon-q |
.codex/config.toml | MCP config or agent settings (servers start from it) | codex |
.gemini/settings.json | MCP config or agent settings (servers start from it) | gemini-cli |
opencode.json | MCP config or agent settings (servers start from it) | opencode |
opencode.jsonc | MCP config or agent settings (servers start from it) | opencode |
.vscode/tasks.json | Editor file that can run a command when the folder opens | vscode-tasks |
.vscode/settings.json | Editor file that can run a command when the folder opens | vscode-settings |
.gemini/.env | Editor file that can run a command when the folder opens | gemini-env |
.codex/hooks.json | Editor file that can run a command when the folder opens | codex-hooks |
.claude/scheduled_tasks.json | Editor file that can run a command when the folder opens | claude-scheduled-tasks |
.claude/workflows | Editor file that can run a command when the folder opens | claude-workflows |
.mcphub/servers.json | Editor file that can run a command when the folder opens | mcphub-config |
.claude/commands/*.md | Read by the agent as instructions | claude-code |
.claude/skills/**/SKILL.md | Read by the agent as instructions | claude-code |
.agents/skills/**/SKILL.md | Read by the agent as instructions | agent-skills |
.gemini/skills/**/SKILL.md | Read by the agent as instructions | gemini-cli |
.cursor/skills/**/SKILL.md | Read by the agent as instructions | cursor |
.codex/skills/**/SKILL.md | Read by the agent as instructions | cursor |
.github/skills/**/SKILL.md | Read by the agent as instructions | copilot |
.claude/CLAUDE.md | Read by the agent as instructions | claude-code |
CLAUDE.md | Read by the agent as instructions | claude-code |
.cursor/rules/*.mdc | Read by the agent as instructions | cursor |
.github/copilot-instructions.md | Read by the agent as instructions | copilot |
.github/instructions/*.instructions.md | Read by the agent as instructions | copilot |
.github/prompts/*.prompt.md | Read by the agent as instructions | copilot |
.claude/agents/**/*.md | Read by the agent as instructions | |
CLAUDE.local.md | Read by the agent as instructions | claude-code |
.claude/rules/**/*.md | Read by the agent as instructions | claude-code |
AGENTS.md | Read by the agent as instructions | agents-md |
.claude/AGENTS.md | Read by the agent as instructions | claude-code |
AGENTS.override.md | Read by the agent as instructions | codex |
SOUL.md | Read by the agent as instructions | openclaw |
USER.md | Read by the agent as instructions | openclaw |
MEMORY.md | Read by the agent as instructions | openclaw |
HEARTBEAT.md | Read by the agent as instructions | openclaw |
.github/workflows/*.yml | GitHub Actions workflow, read for AI agent steps that run on text a stranger can write | github-actions |