← MCP server and skill grades

mcp-toolbox

googleapis/mcp-toolbox

D
Type
MCP server
Scanned
Nov 13, 2026
Version
45ed205cc868
Files read
2052

...` -> wait! No! Do not output extra text or markdown fences! Just the raw JSON string. Let me output only the JSON now.{

Findings

3 high12 medium17 low
  1. HighDefault HTTP server allows cross-origin requests and DNS rebinding from any website

    cmd/internal/flags.go:76

    When run without --stdio (the default CLI and container mode), Toolbox starts an HTTP server on 127.0.0.1:5000 with no authentication unless an MCP authService is explicitly configured. Because --allowed-origins defaults to '*' and --allowed-hosts defaults to '*', any website the user visits in a browser can send cross-origin fetch requests or use DNS rebinding to list and invoke all configured database tools, exfiltrate data, modify databases, or read file resources without AI involvement.

    Fix: Default --allowed-origins to an empty list so cross-origin browser requests are blocked unless explicitly configured, and default --allowed-hosts to localhost and 127.0.0.1 when listening locally.

  2. HighDefault wildcard CORS allows any website to call local HTTP MCP server

    server.json:147

    When MCP Toolbox runs in HTTP mode (the default when --stdio is not used, and the transport configured in server.json), it listens on http://127.0.0.1:5000/mcp without authentication by default and sets --allowed-origins to *. Any malicious website the user visits in a browser on the same computer can send cross-origin POST requests to 127.0.0.1:5000/mcp, pass CORS preflight checks, invoke configured database tools such as execute_sql, and read back the database contents.

    Fix: Do not default --allowed-origins to *; reject cross-origin browser requests unless specific trusted origins are configured, and configure --allowed-hosts for localhost to block DNS rebinding.

  3. HighWildcard allowed-origins default exposes local HTTP MCP server to any website

    server.json:147

    In server.json, the server is configured to run over streamable-http on 127.0.0.1:5000 with --allowed-origins defaulting to *. Because no authentication is enabled by default and CORS permits every origin, any website the user visits in a local browser can send requests to http://127.0.0.1:5000/mcp and read the responses, allowing a malicious page to list tools, run database queries, modify data, or read exposed resources without the AI agent.

    Fix: Restrict the default --allowed-origins value to trusted origins or none, require an authentication token for HTTP requests, and prefer --stdio for local desktop and IDE clients.

  4. MediumFile resourceTemplate prefix bypass via double-slash URI when allowedPaths is omitted

    internal/resources/file/file.go:627

    When a file resourceTemplate defines a directory prefix in uriTemplate (for example, file:///var/log/myapp/{path}) without setting the optional allowedPaths list, requesting a URI with a double slash after the prefix (such as file:///var/log/myapp//Users/alice/tools.yaml) causes MatchResourceTemplateURI to extract an absolute path (/Users/alice/tools.yaml). FileTemplate.Read then skips reconstructing uriTemplate because filepath.IsAbs(pathStr) is true, and checkSandbox only blocks hidden dotfiles when allowedPaths is empty, allowing the caller to read any non-hidden .yaml, .json, .sql, .py, .js, .txt, .md, .csv, or .xml file on the filesystem.

    Fix: Always resolve the extracted path parameter relative to the uriTemplate directory prefix (rejecting leading slashes or drive letters in pathStr), and derive an implicit allowedPaths boundary from the static prefix of uriTemplate when allowedPaths is not set.

  5. MediumWindows command injection in skill wrapper scripts generated by skills-generate

    cmd/internal/skills/generator.go:197

    The skills-generate subcommand defaults to --invocation-mode=npx and emits Node.js wrapper scripts that spawn npx.cmd with shell: true on Windows. Before spawning, arguments containing double quotes are wrapped using '"' + arg.replace(/"/g, '""') + '"'. Because cmd.exe toggles quote state on every '"' character rather than treating '""' as a shell escape, every JSON string value lands outside quotes in cmd.exe (and %VAR% variables expand inside quotes), allowing prompt-injected tool arguments such as {"name":"& calc.exe &"} to execute arbitrary OS commands.

    Fix: Do not pass shell: true to spawn on Windows; instead resolve the npx CLI JavaScript entry point and execute it directly with process.execPath and stdio argument arrays, or use binary invocation without a shell.

  6. MediumMissing backslash escaping in StringParameter escape enables SQL injection in templateParameters

    internal/util/parameters/parameters.go:724

    StringParameter supports escape modes (single-quotes, double-quotes, backticks) intended to make templateParameters safe to interpolate into SQL statements. However, applyEscape only doubles the delimiter character and never escapes backslashes ('\'). In BigQuery (GoogleSQL), MySQL, Cloud SQL MySQL, ClickHouse, TiDB, SingleStore, OceanBase, and MindsDB, '\' escapes the closing quote (and BigQuery and ClickHouse backtick identifiers also use '\' escaping rather than doubled backticks). Passing a value ending in '\' escapes the closing delimiter and allows a second parameter to inject raw SQL outside quotes.

    Fix: Reject or escape backslashes according to the target database dialect, or restrict identifier templateParameters to strict character allowlists rather than generic quote doubling.

  7. MediumCockroachDB readOnlyMode is bypassed by multiline comments, CTEs, UPSERT, and EXPLAIN ANALYZE

    internal/sources/cockroachdb/cockroachdb.go:291

    The cockroachdb source enables readOnlyMode: true by default and relies on ClassifySQL and IsWriteOperation to block writes. ClassifySQL strips block comments without the (?s) flag (failing on any multiline /* ... */ comment) and classifies statements starting with WITH, UPSERT, EXPLAIN ANALYZE, GRANT, or REVOKE as SQLTypeUnknown or SQLTypeExplain, for which IsWriteOperation returns false. A prompt-injected agent can execute arbitrary writes or table drops on a read-only CockroachDB source using multiline comments, writeable CTEs, UPSERT, or EXPLAIN ANALYZE DELETE.

    Fix: Enforce read-only mode at the database session or transaction level (such as default_transaction_read_only = on or read-only transactions) instead of a SQL prefix denylist, and return s.IsReadOnlyMode() from Source.IsReadOnly().

  8. MediumSQL injection in BigQuery forecast and contribution analysis tools bypasses read-only mode

    internal/tools/bigquery/bigqueryforecast/bigqueryforecast.go:204

    Both bigquery-forecast and bigquery-analyze-contribution advertise readOnlyHint: true and remain active when a BigQuery source is configured with readOnly: true (writeMode: blocked) or writeMode: protected. When allowedDatasets is not configured (the default), neither tool dry-runs history_data or input_data to verify it is a single SELECT statement, nor checks BigQueryWriteMode(), before interpolating it raw into (%s). A prompt-injected agent can close the outer parentheses and append semicolon-separated DDL/DML statements (such as DROP TABLE or DELETE) that BigQuery executes as a multi-statement script.

    Fix: Always dry-run both the user-supplied query and the final generated SQL regardless of allowedDatasets, verify StatementType is SELECT (rejecting SCRIPT and write statements), and enforce the source's BigQueryWriteMode().

  9. MediumPath traversal in Cloud Healthcare resource IDs bypasses store and dataset restrictions

    internal/sources/cloudhealthcare/cloud_healthcare.go:486

    The cloud-healthcare source binds tools to a configured project, region, and dataset, and optionally restricts access via allowedFhirStores and allowedDicomStores. While FHIRFetchPage strictly validates full URLs, GetFHIRResource, FHIRPatientEverything, RetrieveRenderedDICOMInstance, and SearchDICOM interpolate agent-supplied IDs (resourceType, resourceID, patientID, StudyInstanceUID, SeriesInstanceUID, SOPInstanceUID, and storeID when allowlists are unset) directly into {+name} URL paths without escaping '/' or rejecting '..'. A prompt-injected agent can traverse into unauthorized FHIR/DICOM stores or datasets.

    Fix: Validate all FHIR and DICOM path segments to reject '/', '\', and '..', and escape each individual segment with url.PathEscape.

  10. MediumRDF mutation injection and misleading readOnlyHint in Dgraph dgraph-dql tool

    internal/sources/dgraph/dgraph.go:420

    When a dgraph-dql tool is configured with isQuery: false to run a parameterized RDF mutation, embedParamsIntoMutation substitutes each parameter using strings.ReplaceAll and fmt.Sprintf('"%v"', value) without escaping double quotes, backslashes, or newlines (and defaults the tool annotation to readOnlyHint: true). A prompt-injected agent can pass double quotes and newlines in a parameter value to break out of the RDF literal and inject arbitrary N-Quad mutations into Dgraph.

    Fix: Escape backslashes, double quotes, and newlines per the N-Quads specification in a single pass over placeholders, and set destructive annotations when isQuery is false.

  11. MediumUnrestricted directory and flag injection in dataform-compile-local with readOnlyHint set

    internal/tools/dataform/dataformcompilelocal/dataformcompilelocal.go:103

    The dataform-compile-local tool marks itself readOnlyHint: true (which MCP clients commonly auto-approve) and passes the agent-supplied project_dir directly to exec.CommandContext(ctx, 'dataform', 'compile', projectDir, '--json') without a '--' separator or any allowed base directory check. Values starting with '-' act as CLI flags, and because dataform compile executes project JavaScript and loads node_modules/@dataform/core from project_dir, pointing it at an attacker-controlled or downloaded directory executes arbitrary code.

    Fix: Pin or restrict project_dir to an operator-configured allowed directory in tools.yaml, reject arguments starting with '-' (or pass '--' before projectDir), and set readOnlyHint: false.

  12. MediumNeo4j read-only Cypher classifier can be bypassed to execute write and delete queries

    internal/tools/neo4j/neo4jexecutecypher/classifier/classifier.go:209

    When readOnly is enabled for Neo4j Cypher execution, QueryClassifier.Classify attempts to block write operations using regular expressions. However, normalizeQuery strips /* ... */ and // comments before stringLiteralPattern removes string literals, so a valid Cypher statement such as WITH '/*' AS a MATCH (n) DETACH DELETE n RETURN '*/' hides DETACH DELETE inside what the classifier treats as a comment while Neo4j executes the deletion. The classifier can also be bypassed using escaped quotes inside strings ('a\'b'), backtick-quoted identifiers or procedure names without spaces (such as DELETEn or CALL apoc.create.node(...)), and LOAD CSV statements (which can trigger outbound HTTP requests or read files).

    Fix: Enforce read-only mode at the Neo4j driver/session level using neo4j.AccessModeRead (and a read-only database role) rather than relying on regex-based Cypher query classification.

  13. MediumHTTP MCP server allows any browser origin and host by default

    server.json:147

    When MCP Toolbox runs in its default HTTP mode (including the streamable-http package configuration in server.json), --allowed-origins defaults to * and --allowed-hosts is omitted (also defaulting to *), with no authentication required unless configured in tools.yaml. Because the local server on 127.0.0.1:5000 returns Access-Control-Allow-Origin: * and accepts any Host header, any malicious website the user visits in a browser can directly send requests to /mcp via cross-origin fetch or DNS rebinding to list tools, read resources, and run database queries with no AI agent involved.

    Fix: Default --allowed-origins to an empty list so cross-origin browser requests are blocked unless explicitly configured, expose and default --allowed-hosts to 127.0.0.1,localhost, or default local MCP client configurations to --stdio.

  14. MediumPR review skill instructs agent to check out and run untrusted fork PR code

    skills/maintainer/review-prs/SKILL.md:210

    The review-prs Agent Skill instructs the coding agent to prefer running code over reasoning about it by fetching an untrusted pull request branch (pull/<n>/head), checking it out into a worktree, adding a scratch probe_test.go inside the package under review, and running go test. Anyone on GitHub can open a pull request against a public repository, and running go test on their branch executes package init() functions, cgo directives, modified code, and go.mod replacements on the reviewer's machine with access to their GitHub token and cloud credentials, bypassing CI protections that gate fork PRs.

    Fix: Remove the instruction to run go test on unvetted pull/<n>/head branches locally, or require explicit user approval after auditing the full diff (including init() functions, go.mod, build tags, and cgo directives) and run any probe inside an isolated, network-disabled container without host credentials.

  15. MediumPR review skill instructs coding agent to fetch and run untrusted PR code

    skills/maintainer/review-prs/SKILL.md:210

    Although the review-prs skill describes itself as PROPOSE-ONLY, its Rules section instructs the AI agent to prefer running the pull request's code over reasoning about it by fetching pull/<n>/head into a worktree and running a scratch probe_test.go inside the package under review. Reviewing a pull request from an external stranger and running go test on their branch executes any attacker-controlled init() functions, #cgo directives, or modified dependencies on the reviewer's machine, where authenticated gh and cloud credentials are present.

    Fix: Remove the instruction to run go test on unreviewed pull/<n>/head branches locally, or restrict probes to main (HEAD) and isolated, credential-free sandboxes after all changed files have been vetted.

  16. LowURL path injection via unescaped projectId in Cloud Monitoring RunQuery

    internal/sources/cloudmonitoring/cloud_monitoring.go:142

    RunQuery interpolates the agent-provided projectId parameter directly into the Cloud Monitoring URL without validation or url.PathEscape while using credentials scoped to monitoring.MonitoringScope. Supplying a '#' fragment (and/or '/' or '..' segments, such as '../../v3/projects/my-proj/notificationChannels#') truncates the Prometheus query suffix into the URL fragment and sends an authenticated GET request to arbitrary Cloud Monitoring API endpoints, potentially exposing sensitive notification channel tokens or configs.

    Fix: Validate projectID against GCP project ID naming rules and apply url.PathEscape(projectID) when constructing the request URL.

  17. LowNon-MCP Google authService accepts tokens for any audience when clientId is omitted

    internal/auth/google/google.go:126

    In google.Config, ClientID has no required struct tag, and Config.Initialize only checks that audience or clientId is non-empty when mcpEnabled is true. If an operator configures a google authService with mcpEnabled: false (the default) and omits clientId, GetClaimsFromHeader calls idtoken.Validate(ctx, token, '') with an empty audience string, causing google.golang.org/api/idtoken to skip aud claim verification and accept valid Google ID tokens issued to any unrelated third-party OAuth application.

    Fix: Require clientId to be non-empty in Config.Initialize when mcpEnabled is false, and return an error in GetClaimsFromHeader if a.ClientID is empty.

  18. LowMaintainer review-prs skill instructs agent to run go test on unvetted external pull requests

    skills/maintainer/review-prs/SKILL.md:209

    The shipped review-prs Agent Skill instructs maintainers' coding agents to fetch external pull request branches ('git fetch origin pull/<n>/head:pr<n>'), check them out into a worktree, and run Go tests inside the package under review. Running go test on an unvetted PR executes attacker-controlled init() functions, test code, or modified dependencies on the reviewer's machine with access to local cloud and GitHub credentials.

    Fix: Update the skill instructions to forbid building or running code from external PRs until the entire diff (including init() functions, go.mod, CGO directives, and scripts) has been audited for malicious code, or require running probes inside an isolated container.

  19. LowSQL injection in prebuilt SQLite list_tables tool via unescaped template parameters

    internal/prebuiltconfigs/tools/sqlite.yaml:101

    The prebuilt SQLite list_tables tool interpolates the agent-supplied output_format and table_names arguments directly into single-quoted SQL literals using templateParameters instead of bound SQLite ? parameters. Because the modernc.org/sqlite driver executes multiple semicolon-separated statements in one call, a prompt-injected agent calling list_tables can break out of the quote and run arbitrary SQL statements—such as DROP TABLE, DELETE, or ATTACH DATABASE to create or overwrite files on disk—through a tool that appears to only inspect table schemas.

    Fix: Replace templateParameters with standard SQLite positional ? parameters in list_tables (matching the postgres-list-tables and mysql-list-tables implementations) and restrict output_format with allowedValues.

  20. LowConnected-schema restriction in MySQL monitoring tools can be bypassed via the connected_schema parameter

    internal/tools/mysql/mysqlshowquerystats/mysqlshowquerystats.go:139

    mysql-show-query-stats, mysql-list-all-locks, and mysql-list-table-stats attempt to restrict queries to the configured source database (source.MySQLDatabase()) by rejecting calls where table_schema differs from connected_schema. However, connected_schema is exposed as an ordinary tool parameter that the AI agent can set directly, and source.MySQLDatabase() is only used when connected_schema is empty. A prompt-injected agent can pass the same target database name in both table_schema and connected_schema to bypass this check and read query digests, active transaction SQL snippets, and table statistics from other schemas on the MySQL server.

    Fix: Remove the connected_schema parameter from the tool manifest and always compare table_schema directly against source.MySQLDatabase().

  21. LowUnescaped query interpolation in prebuilt PostgreSQL and AlloyDB get_query_plan tools

    internal/prebuiltconfigs/tools/alloydb-omni.yaml:193

    The prebuilt get_query_plan tool in alloydb-omni.yaml, alloydb-postgres.yaml, cloud-sql-postgres.yaml, and postgres.yaml interpolates {{.query}} raw into EXPLAIN (FORMAT JSON) {{.query}}; and includes the tool in read-only monitor groups where execute_sql is absent. If simple query protocol is enabled via queryParams (such as default_query_exec_mode=simple_protocol for PgBouncer compatibility), a semicolon in query injects and executes arbitrary SQL statements. While postgres.yaml warns against production use, alloydb-omni.yaml omits the warning and alloydb-postgres.yaml explicitly tells users to use it safely in production.

    Fix: Implement a dedicated Go tool for PostgreSQL EXPLAIN that rejects multiple statements, executes inside a READ ONLY transaction that is rolled back, and remove misleading production-safety claims.

  22. LowSQL injection in prebuilt MindsDB parameterized_sql tool via unescaped template parameters

    internal/prebuiltconfigs/tools/mindsdb.yaml:39

    The prebuilt MindsDB parameterized_sql tool describes itself as executing parameterized SQL queries, but interpolates table_name and condition_column directly into SELECT * FROM {{.table_name}} WHERE {{.condition_column}} = ? LIMIT {{.limit}} without identifier validation or escaping. A prompt-injected agent can pass arbitrary SQL clauses (such as UNION SELECT or subqueries) in table_name or condition_column.

    Fix: Validate table_name and condition_column as strict SQL identifiers or quote and escape them before interpolating into the SQL statement.

  23. LowCypher injection via unescaped database labels and relationship types in neo4j-schema

    internal/tools/neo4j/neo4jschema/neo4jschema.go:368

    When APOC procedures are unavailable, GetSchemaWithoutAPOC retrieves node labels and relationship types from Neo4j and interpolates them directly into MATCH (n:%s) and MATCH ()-[r:%s]->() without backtick quoting or escaping, running on a session that does not set neo4j.AccessModeRead. If untrusted graph data introduced a crafted label or relationship type, invoking the read-only neo4j-schema tool executes injected Cypher Clauses.

    Fix: Wrap labels and relationship types in backticks with doubled-backtick escaping (as done in falkordbschema) and open schema-inspection sessions with neo4j.AccessModeRead.

  24. LowMissing collection path validation after template substitution in firestore-query

    internal/tools/firestore/firestorequery/firestorequery.go:177

    Unlike the other Firestore tools in this repository, firestore-query populates its collectionPath template from agent parameters without calling fsUtil.ValidateCollectionPath or blocking '/' and '..' in parameter values. In a custom tool intended to scope queries to a specific subcollection template such as users/{{.userId}}/orders, a prompt-injected agent can pass path separators or '..' segments to query unintended collections.

    Fix: Reject '/' and '..' in template parameter values used to build collectionPath and call fsUtil.ValidateCollectionPath(collectionPath) before building the query.

  25. LowUnvalidated path segments in looker-generate-embed-url allow path traversal outside /embed/

    internal/tools/looker/lookergenerateembedurl/lookergenerateembedurl.go:150

    looker-generate-embed-url builds TargetUrl with fmt.Sprintf("%s/embed/%s/%s", hostURL, embedType, contentId) and calls CreateEmbedUrlAsMe to mint a signed, authenticated Looker session URL. Neither embedType nor contentId is restricted to an allowlist, URL-escaped, or checked for '..' traversal, '?' query strings, or '#' fragments, allowing a prompt-injected agent to request signed login URLs targeting arbitrary paths on the Looker host.

    Fix: Restrict type to an explicit allowlist (such as dashboards, looks, query-visualizations, explores), reject '..' and query/fragment delimiters, and URL-escape path segments.

  26. LowMaintainer review-prs skill instructs the agent to fetch and execute code from untrusted pull requests

    skills/maintainer/review-prs/SKILL.md:211

    The installable review-prs skill instructs coding agents to prefer running code over reasoning about it by fetching pull/<n>/head into a temporary git worktree and executing a scratch probe_test.go inside the package under review. Running go test on an unvetted external fork PR executes attacker-controlled init() functions, modified package code, and #cgo directives directly on the reviewer's machine with access to local credentials.

    Fix: Instruct the agent never to run go test, go build, or repository scripts on code from an external fork PR until the entire diff (including init functions, go.mod, and cgo directives) has been inspected and the user explicitly confirms local execution.

  27. Lowwait tool ignores its required timeout configuration and context cancellation

    internal/tools/utility/wait/wait.go:95

    The wait tool requires a timeout field in its YAML configuration, but Invoke never reads or enforces t.Cfg.Timeout and calls time.Sleep(totalDuration) directly instead of waiting on ctx.Done(). A prompt-injected agent can supply an arbitrarily large duration (such as 2500000h) to block goroutines indefinitely even after client cancellation.

    Fix: Parse and enforce t.Cfg.Timeout in Initialize/Invoke to cap totalDuration, and use time.NewTimer with a select on ctx.Done() so cancelled requests return immediately.

  28. LowWrite-capable tools default to readOnlyHint=true annotations

    internal/tools/mongodb/mongodbaggregate/mongodbaggregate.go:97

    mongodb-aggregate (when readOnly is false, its default, allowing $out and $merge writes), neo4j-cypher, falkordb-cypher, and looker-dev-mode (which mutates the Looker workspace session) default to readOnlyHint=true and destructiveHint=false in their MCP tool annotations. MCP clients that auto-approve tools marked readOnlyHint=true may let a prompt-injected agent invoke these state-changing tools without user confirmation.

    Fix: Default to destructive/write annotations whenever readOnly is false (as spanner-execute-sql and falkordb-execute-cypher do), and use write annotations on looker-dev-mode.

  29. LowBug reproduction skill instructs agent to pass untrusted issue YAML directly to toolbox

    skills/maintainer/reproduce-bug/SKILL.md:127

    The reproduce-bug Agent Skill instructs the agent to feed a reporter's YAML configuration from a GitHub issue directly to the local toolbox binary to check config parsing, assuming config validation runs before any connection is made. If an issue reporter supplies a syntactically valid tools.yaml, toolbox immediately expands ${ENV_VAR} placeholders from the local environment and initializes configured sources, which can connect to attacker-controlled database hosts or custom API endpoints and leak local environment secrets or Google Cloud credentials.

    Fix: Instruct the agent to strip or replace all ${ENV_VAR} interpolations, external hosts and URIs, custom API endpoints, and local file paths in reporter-supplied YAML before running toolbox, or run reproductions inside an isolated environment without credentials or network access.

  30. LowBug reproduction skill feeds untrusted issue YAML directly to toolbox binary

    skills/maintainer/reproduce-bug/SKILL.md:127

    Rung 1 of the reproduce-bug skill instructs the AI agent to feed an issue reporter's YAML config directly to the toolbox binary to check startup errors, assuming parsing happens with zero infrastructure. If a stranger's YAML in a GitHub issue is syntactically valid, toolbox expands ${ENV_NAME} environment variables from the local shell and immediately connects to configured sources at startup, which can send local secrets (such as GITHUB_TOKEN or API keys) to an attacker-controlled host, use local Application Default Credentials, or create local SQLite files.

    Fix: Instruct the agent to strip all ${...} environment variable expansions, external host or URI endpoints, cloud project IDs, and outside file paths from reporter YAML configs before passing them to toolbox.

  31. LowMaintainer PR review skill instructs agent to execute untrusted pull request code

    skills/maintainer/review-prs/SKILL.md:211

    Although the optional review-prs maintainer skill describes itself as PROPOSE-ONLY, its rules tell the coding agent to prefer running the code by fetching pull/<n>/head into a /tmp/pr<n> worktree and running a scratch probe_test.go inside the package under review. If a maintainer uses this skill on an untrusted external pull request, running go test executes attacker-controlled init() functions, test helpers, or modified go.mod/cgo directives with the maintainer's local credentials (rated low as maintainer-only tooling not run by the installed MCP server).

    Fix: Never run go test on an unreviewed pull request branch on the host machine; require human inspection of the full diff first or execute probes only inside an isolated container without network or credential access.

  32. LowMaintainer bug reproduction skill feeds untrusted issue YAML directly to Toolbox

    skills/maintainer/reproduce-bug/SKILL.md:127

    The optional reproduce-bug maintainer skill instructs the coding agent to take a bug reporter's YAML config from a GitHub issue and feed it directly to the toolbox binary to check startup errors, assuming config parsing runs before any connection. Because Toolbox expands ${ENV_VAR} placeholders in YAML files and initializes configured sources on startup, an attacker-crafted issue config could exfiltrate local environment variables or access local files when run on a maintainer's machine (rated low as maintainer-only tooling not run by the installed MCP server).

    Fix: Require the agent to strip ${...} environment variable interpolations, local file paths, and external hostnames or URLs from reporter-supplied YAML configs before passing them to the toolbox binary.