MCP06
- Location
entry.mjs:811- Finding
Remote Service Messaging Is Endorsed as Agent Instructions ## Remediation Suggestions 1. **Remove instructional endorsement of remote text.** Identify service messaging as untrusted display content. Do not tell the agent to follow it or relay it verbatim. 2. **Generate workflow instructions locally.** Validate structured status values and map them to fixed, locally authored guidance. Remote messaging must not determine tool selection, security policy, or access to unrelated resources. 3. **Constrain messaging schemas.** Accept only expected fields with explicit types and length limits. Reject unexpected instruction-bearing fields rather than forwarding arbitrary objects. 4. **Validate account-claim links.** Require HTTPS, an operator-approved service origin, and an expected claim-path structure before presenting a link as an account-claim action. 5. **Apply a consistent external-content boundary.** Cover registration, matchmaking, tournament, game-over, rules, and strategy responses. Explicitly state that remote content cannot authorize credential disclosure, filesystem access, unrelated tool calls, or policy changes. 6. **Enforce sensitive actions outside the model.** Use gateway permissions and approval requirements to prevent remote-content-driven access to secrets or unrelated tools. 7. **Add adversarial response tests.** Test remote messages containing tool-use instructions, secret requests, and malicious claim URLs. Verify that tool results retain untrusted-data framing and do not endorse those requests.
- Content
View full analysis
Vulnerability Details
File Location:
entry.mjs:811–814; supporting input flow attransport.mjs:222–224
Vulnerability Type: Indirect Prompt Injection / External Content Trust-Boundary Violation
Risk Level: MediumComplete Code Snippet
The HTTP transport preserves the service-provided messaging object:
js ...(res.data.messaging && typeof res.data.messaging === 'object' ? { messaging: res.data.messaging } : {}),The registration handler forwards that messaging into the model-visible result and explicitly endorses it:
js ...(r.messaging && typeof r.messaging === 'object' ? { messaging: r.messaging } : {}), guidance: 'Relay messaging.forOperator to your operator so they can claim this agent; messaging.toAgent is written for you. The plugin has already stored your API key and sends it on every call, so you never need it.',Technical Analysis
The registration request receives JSON from the configured SteamedClaw service.
client.register()copies the response’smessagingobject, and the registration tool includes it in its JSON text result.The plugin’s locally authored guidance tells the model that
messaging.toAgentis written for it and directs it to relaymessaging.forOperator. This promotes externally supplied presentation content into an instruction-bearing surface. The same general pattern appears in matchmaking, tournament, and game-over messaging.This is distinct from ordinary game-data passthrough: the plugin actively endorses the remote text instead of consistently identifying it as untrusted data. Opponent chat receives an explicit non-instruction boundary elsewhere in the plugin, but this messaging surface does not.
The confirmed issue is the unsafe trust relationship. Successful behavioral exploitation remains conditional on attacker control of upstream messaging and the model following the injected content. The reviewed artifact does not establish that ordinary players can modify these mess ...[truncated 1645 chars]
