ShimShield ShimShield  ·  Attack Prevention Shim
Open-access security research  /  MITRE ATLAS™ contributions

The agent portcullis works both ways.

ShimShield mediates the wire between an AI agent and the world in both directions: it decides whether an agent may operate on a channel at all, and what an admitted agent is allowed to do. Each direction is a reproducible, cryptographically verifiable contribution to MITRE ATLAS. Read either white paper below.

Enforcement: at the wire Evidence: tamper-evident SHA-256 chains Provenance: Bitcoin mainnet Reproducible: one command each
Two directions, one primitive

Choose a direction

Outbound governs the agent's tool calls; inbound governs the agent's admission. Together they close the boundary.

Outbound traffic

Wire-Level Tool-Call Mediation

Once an agent is admitted, what may it do? Every tool call is authorized against a signed allow-list at the transport, refused if not permitted, and recorded, independent of the agent's reasoning.

3 mitigations 6 / 6 case studies reproduced on-chain anchor
T0053 T0086 T0101 T0098 T0110 T0109 T0080 …
Read the white paper →
Inbound traffic

Wire-Level Agent Admission (Shimwall)

Before any of that, may this agent operate here at all? An agent is admitted only on cryptographic proof of an authorized identity, by signing a fresh challenge, refusing the unauthorized, the forged, and the replayed at the wire.

1 mitigation 1 / 1 case study reproduced on-chain anchor
T0103 T0108 T0073 T0091
Read the white paper →
Every claim on both papers is designed to be re-run, not believed: each case study reproduces with a single command against the shipping code, seals its verdicts in a tamper-evident SHA-256 hash chain a reviewer re-verifies, and anchors its provenance on the Bitcoin main chain. Coverage is claimed only where ShimShield enforces at the wire.