Block hijacked tool calls after prompt injection. Lineage Firewall and mandatory hop — not keyword matching alone.

The problem

‍

Injection is a session problem, not a string problem. Blocking DROP TABLE in one message misses the attack. Contaminated lineage means the next authorized-looking tool call is the exfiltration or destructive action.

‍

‍

How Blekline handles it

‍

  • Lineage Firewall — block destructive tools on contaminated sessions.
  • Mandatory hop — every tool call evaluated at the sidecar before downstream execution.
  • Contracts enforce — @blekline/contracts evaluates tool arguments, not just prompts.
  • Runtime simulator — contaminate → block demo for eval and board conversations.
  • MCP supply-chain guard — approved downstream server registry (enhanced in enterprise program).

‍

‍

Mandatory hop pattern

‍

Agent → POST /v1/enforce-tool-call → (allow) → downstream tool / MCP ↑ blekline-sidecar

‍

‍

What ships today

‍

ControlWhereStatusLineage block (demo)Runtime simulator + sidecarShipped v1Tool-arg enforceMCP proxy + contractsShippedFleet kill switchSidecar + SSEShippedSPIFFE / IAM attestationSidecar roadmapSimulated in demo

‍

‍

Try it

‍

Open the Runtime Simulator — pick Prompt injection → DROP TABLE, execute, and read the trace in the interceptor console.

‍

‍

Honest scope

‍

Simulator proves policy UX and contracts behavior. Production lineage attestation and SPIFFE binding require Track 01/02 sidecar deployment — not playground-only evidence for audit packets.

Healthcare AI security use case

Qualified for fleet rollout?