prompt-injection-threat-model

Assess indirect prompt injection risk by mapping untrusted text exposure across every input channel in your Claude harness, including user prompts, tool output, web content, and MCP server responses. Perform channel-by-channel trust-boundary analysis to identify where injections could succeed.

A prompt injection threat model maps how untrusted content from MCP servers, hooks, and subagent outputs crosses trust boundaries to influence Claude. It evaluates exposure and blast radius for each channel to determine long-term hardening needs against indirect prompt injection.

Harden Claude by reviewing current defenses and applying mitigation patterns like content labeling, privilege separation, and confirmation gates. Create a practical hardening plan that reduces injection risk from web content and files without breaking useful workflows.

Yes, the analysis maps prompt-injection exposure across user prompts, tool output, web content, MCP server responses, subagent output, hook output, and memory injection. It assesses the trust boundary and blast radius for each specific channel in your setup.

The best way to review MCP security is to conduct a channel-by-channel trust-boundary analysis and blast-radius assessment. This identifies indirect prompt injection risks in MCP server responses and produces a concrete mitigation plan grounded in the actual tools and models in use.

Hardening limitations depend on the actual tools and models in use, requiring trade-offs to avoid breaking useful workflows. Mitigation plans must balance near-term defenses like confirmation gates with long-term privilege separation across untrusted input channels.