⌘K

Findings

What a finding is, what it contains, and the categories MCPSafe detects.

Finding object

{
  "finding_id": "f_001",
  "mcp_threat_id": "MCP-205",
  "category": "injection",
  "severity": "high",
  "aivss_score": 7.4,
  "tool_name": "execute_command",
  "file_path": "src/tools/execute.ts",
  "line_number": 42,
  "description": "User input flows into an inner-LLM SYSTEM message",
  "remediation": "Wrap untrusted input with <untrusted>...</untrusted> or pass through escape_for_prompt()",
  "evidence_snippet": "messages: [{ role: 'system', content: `Use this context: ${userInput}` }]"
}

Field	Description
`finding_id`	Stable identifier within this scan
`mcp_threat_id`	The rule that produced the finding (e.g. `MCP-205`); cross-references /threats/coverage
`category`	Signal category (see below)
`severity`	`critical` / `high` / `medium` / `low`
`aivss_score`	0–10 score for this specific finding
`tool_name`	The MCP tool where the issue was found

Finding categories

Categories are signal buckets, not individual rules. Multiple rules can map to the same category — see /threats/coverage for the active rule set per category.

Category	What it covers
`injection`	Prompt injection into inner LLMs, command/SQL injection through tool input
`secrets`	Runtime secret exfiltration to logs/responses, plaintext secrets in `.env`
`permissions`	IAM wildcards, OAuth over-scoping, unused-scope detection
`supply_chain`	Typosquats, install-time remote-exec hooks, container running as root, known CVEs
`destructive`	Mutating tool handlers without `elicit()` / `dry_run` / confirmation
`cve`	NVD / GHSA matches against the resolved version
`typosquat`	Look-alike package names targeting popular packages

How findings are generated

Fast scans run pattern, manifest, and supply-chain checks across the full catalog.

Deep scans add taint and AI-assisted semantic checks plus an LLM judge panel that votes on each tool handler. The judge panel is a second opinion — it adjusts the score, but rule findings drive the verdict.

New rules go through a precision review before they affect any user-visible result.

Filtering findings

On the scan result page, use the severity and category filters to narrow the findings list. You can also filter programmatically via the API response:

const critical = result.findings.filter(f => f.severity === "critical");
const injections = result.findings.filter(f => f.category === "injection");
const fromRule  = result.findings.filter(f => f.mcp_threat_id === "MCP-205");