mcp-tool-output-hygiene-v1
9 rules · Keryn Software Licence · rules 0.8.0 · all packs
How to read a rule
- Severity
- How bad the finding is if it is real:
errorfails a default gate,warningandinfodo not. - Confidence
- P0 exploitable as written; P1 dangerous in context; P2 a heuristic or a posture check: confirm by reading the code.
- False-positive risk
- low: an exact API or flag, so hits are almost always real; medium: the pattern is legitimate in many programs, so check each hit; high: it finds a lead to confirm, so expect benign hits.
- Edge
- The relationship the finding sits on. reaches: content or traffic gets to a thing; calls: one thing invokes another; reads: one thing reads from another; writes: one thing writes to another; sends: data leaves for another party; grants: authority is given or left unchecked; depends_on: something relies on a component it does not verify; trusts: one party believes another without evidence. The boundary it crosses, when there is one: user->prompt (untrusted input reaching a prompt); model->code (model output reaching code, a shell, a query or a page); secret->outside (a secret leaving the trusted side); supply-chain (code, models or prompts pulled in from outside).
- Taxonomy chips
- References for readers who work in OWASP LLM 2025, OWASP Agentic, CWE or NIST AI RMF. They say where a rule belongs, not that a framework is satisfied.
keryn.mcp.py.tool-returns-credential-like-keys
Section titled “keryn.mcp.py.tool-returns-credential-like-keys”warning python mcp, python, fastmcp
An MCP tool returns a dict whose key names a credential (api_key, password, token, secret, …) with a non-literal value; everything a tool returns lands in the model context, the client transcript and logs. Drop the field or return a redacted form such as the last four characters.
Confidence P2: a heuristic or a posture check: confirm by reading the code. False-positive risk medium: the pattern is legitimate in many programs, so check each hit.
Edge inspected: grants authority is given or left unchecked.
CWE-200 CWE-532 LLM02 LLM06 ASI03
Fix: Do not put credentials in tool results. Return only the data the model needs; if a credential must be confirmed, return a boolean such as configured: true or a masked suffix.
References: https://modelcontextprotocol.io/specification/2025-06-18/server/tools · https://github.com/modelcontextprotocol/python-sdk
keryn.mcp.ts.tool-returns-credential-like-keys
Section titled “keryn.mcp.ts.tool-returns-credential-like-keys”warning typescript, javascript mcp, typescript, nodejs
An MCP tool handler returns an object whose key names a credential (apiKey, password, token, secret, …) with a non-literal value; tool results flow into the model context, the client transcript and logs. Remove the field or return a masked or boolean form.
Confidence P2: a heuristic or a posture check: confirm by reading the code. False-positive risk medium: the pattern is legitimate in many programs, so check each hit.
Edge inspected: grants authority is given or left unchecked.
CWE-200 CWE-532 LLM02 LLM06 ASI03
Fix: Do not put credentials in tool results. Return only the data the model needs; to confirm a credential exists, return a boolean or a masked suffix.
References: https://modelcontextprotocol.io/specification/2025-06-18/server/tools · https://github.com/modelcontextprotocol/typescript-sdk
keryn.mcp.py.tool-returns-fetched-content-unwrapped
Section titled “keryn.mcp.py.tool-returns-fetched-content-unwrapped”warning python data-flow rule mcp, python, fastmcp, requests, httpx
An MCP tool returns the body of an outbound HTTP response to the client verbatim; whoever controls that page can plant instructions that the client model reads as tool output (indirect prompt injection). Wrap the content in an explicit untrusted-data delimiter and strip markup before returning it.
Confidence P2: a heuristic or a posture check: confirm by reading the code. False-positive risk medium: the pattern is legitimate in many programs, so check each hit.
Edge inspected: reaches content or traffic gets to a thing; it crosses the user->prompt boundary (untrusted input reaching a prompt).
CWE-74 CWE-1427 LLM01 LLM05 ASI01
Fix: Treat fetched pages as data: strip HTML, truncate, and wrap the text in a clear untrusted-content envelope (for example <untrusted_content source=URL>…</untrusted_content>) that the client prompt tells the model not to obey. Prefer returning extracted fields over raw bodies.
References: https://modelcontextprotocol.io/specification/2025-06-18/server/tools · https://github.com/modelcontextprotocol/python-sdk · https://modelcontextprotocol.io/docs/2025-11-25/tutorials/security/security_best_practices
keryn.mcp.ts.tool-returns-fetched-content-unwrapped
Section titled “keryn.mcp.ts.tool-returns-fetched-content-unwrapped”warning typescript, javascript data-flow rule mcp, typescript, nodejs
An MCP tool handler returns the body of an outbound HTTP response (fetch or axios) verbatim; whoever controls that page can plant instructions that the client model reads as tool output (indirect prompt injection). Wrap it in an untrusted-data envelope and strip markup first.
Confidence P2: a heuristic or a posture check: confirm by reading the code. False-positive risk medium: the pattern is legitimate in many programs, so check each hit.
Edge inspected: reaches content or traffic gets to a thing; it crosses the user->prompt boundary (untrusted input reaching a prompt).
CWE-74 CWE-1427 LLM01 LLM05 ASI01
Fix: Treat fetched pages as data: strip HTML, truncate, and wrap the text in a clear untrusted-content envelope that the client prompt tells the model not to obey. Prefer returning extracted fields over raw bodies.
References: https://modelcontextprotocol.io/specification/2025-06-18/server/tools · https://github.com/modelcontextprotocol/typescript-sdk · https://modelcontextprotocol.io/docs/2025-11-25/tutorials/security/security_best_practices
keryn.mcp.py.client-plain-http-transport
Section titled “keryn.mcp.py.client-plain-http-transport”warning python mcp, python, client
An MCP client transport connects to a non-loopback http:// URL, so tool calls, results and any bearer token cross the network unencrypted and can be read or rewritten in transit. Use https://; keep http only for localhost.
Confidence P1: dangerous in context. False-positive risk low: an exact API or flag, so hits are almost always real.
Edge inspected: sends data leaves for another party; it crosses the secret->outside boundary (a secret leaving the trusted side).
CWE-319 CWE-300 LLM02 LLM03 ASI07 ASI04
Fix: Point the transport at an https:// endpoint with a valid certificate. Plain http is acceptable only for loopback development servers.
References: https://modelcontextprotocol.io/specification/2025-06-18/basic/transports · https://modelcontextprotocol.io/specification/2025-06-18/basic/authorization · https://github.com/modelcontextprotocol/python-sdk · https://pypi.org/project/mcp/1.12.0/ · https://raw.githubusercontent.com/modelcontextprotocol/python-sdk/main/src/mcp/client/streamable_http.py
keryn.mcp.ts.client-plain-http-transport
Section titled “keryn.mcp.ts.client-plain-http-transport”warning typescript, javascript mcp, typescript, nodejs, client
An MCP client transport connects to a non-loopback http:// URL, so tool calls, results and any bearer token cross the network unencrypted and can be read or rewritten in transit. Use https://; keep http only for localhost.
Confidence P1: dangerous in context. False-positive risk low: an exact API or flag, so hits are almost always real.
Edge inspected: sends data leaves for another party; it crosses the secret->outside boundary (a secret leaving the trusted side).
CWE-319 CWE-300 LLM02 LLM03 ASI07 ASI04
Fix: Point the transport at an https:// endpoint with a valid certificate. Plain http is acceptable only for loopback development servers.
References: https://modelcontextprotocol.io/specification/2025-06-18/basic/transports · https://modelcontextprotocol.io/specification/2025-06-18/basic/authorization · https://github.com/modelcontextprotocol/typescript-sdk · https://raw.githubusercontent.com/modelcontextprotocol/typescript-sdk/main/packages/client/src/client/streamableHttp.ts
keryn.mcp.go.sampling-request-no-approval
Section titled “keryn.mcp.go.sampling-request-no-approval”warning go mcp, go
A Go MCP tool handler asks the client to run a model completion (RequestSampling / CreateMessage) with no approval, consent or policy call anywhere in the same function. Add an explicit gate and token cap before spending the client’s model.
Confidence P2: a heuristic or a posture check: confirm by reading the code. False-positive risk high: it finds a lead to confirm, so expect benign hits.
Edge inspected: grants authority is given or left unchecked.
CWE-862 CWE-345 LLM06 LLM01 ASI09 ASI02
Fix: Sampling lets a server spend the client’s model and credentials. Gate each request behind an explicit approval or policy check (per-server allow-list, token cap, rate limit) and never build the sampled prompt from untrusted tool input without review.
References: https://modelcontextprotocol.io/specification/2025-06-18/client/sampling · https://github.com/mark3labs/mcp-go · https://github.com/modelcontextprotocol/go-sdk · https://raw.githubusercontent.com/modelcontextprotocol/go-sdk/main/mcp/server.go · https://raw.githubusercontent.com/mark3labs/mcp-go/main/server/sampling.go
keryn.mcp.java.sampling-request-no-approval
Section titled “keryn.mcp.java.sampling-request-no-approval”warning java mcp, java
A Java MCP tool handler calls exchange.createMessage(…) to run a completion on the client with no approval, consent or policy call in the same handler. Add an explicit gate and a token cap before spending the client’s model.
Confidence P2: a heuristic or a posture check: confirm by reading the code. False-positive risk high: it finds a lead to confirm, so expect benign hits.
Edge inspected: grants authority is given or left unchecked.
CWE-862 CWE-345 LLM06 LLM01 ASI09 ASI02
Fix: Sampling lets a server spend the client’s model and credentials. Gate each request behind an explicit approval or policy check (per-server allow-list, token cap, rate limit) and never build the sampled prompt from untrusted tool input without review.
References: https://modelcontextprotocol.io/specification/2025-06-18/client/sampling · https://github.com/modelcontextprotocol/java-sdk · https://raw.githubusercontent.com/modelcontextprotocol/java-sdk/main/mcp-core/src/main/java/io/modelcontextprotocol/server/McpSyncServerExchange.java
keryn.mcp.csharp.sampling-request-no-approval
Section titled “keryn.mcp.csharp.sampling-request-no-approval”warning csharp mcp, csharp, dotnet
An [McpServerTool] method asks the client to run a completion (SampleAsync / RequestSamplingAsync) with no approval, consent or policy call in the same method. Add an explicit gate and a token cap before spending the client’s model.
Confidence P2: a heuristic or a posture check: confirm by reading the code. False-positive risk high: it finds a lead to confirm, so expect benign hits.
Edge inspected: grants authority is given or left unchecked.
CWE-862 CWE-345 LLM06 LLM01 ASI09 ASI02
Fix: Sampling lets a server spend the client’s model and credentials. Gate each request behind an explicit approval or policy check (per-server allow-list, token cap, rate limit) and never build the sampled prompt from untrusted tool input without review.
References: https://modelcontextprotocol.io/specification/2025-06-18/client/sampling · https://github.com/modelcontextprotocol/csharp-sdk · https://raw.githubusercontent.com/mark3labs/mcp-go/main/server/sampling.go