# Evolution 85 — Threat Model

**FIRME AI CORPORATION · CAIN-42 · Evolution 85** — branch `evolution-85`, baseline Evolution 84 (`2c419a042`). Development result. Not deployed. Not released. Built by the three-person team.

## Assets

Authority to act (grants), the resource's state, the kill-switch state, the receipt log, trusted knowledge, the
evaluation's held-out set, signing keys.

## Adversaries and results

| Adversary | Attempt | Control | Result |
|---|---|---|---|
| Compromised or over-eager agent | act without delegation, widen scope, use another agent's chain | delegation chain verification | refused |
| Agent with a valid grant | change the arguments after authorization | exact-request digest at enforcement | refused |
| Agent replaying | reuse a grant or an enforcement record | single-use stores at enforcement and at the resource | refused |
| Agent after shutdown | use a cached grant, retry, delayed message | kill epochs bound into grants | refused, also after reinstate |
| Caller bypassing the boundary | call the resource with no or forged enforcement evidence | resource-side verification of a pinned key | refused |
| Same-host process | write the resource file directly | seal audit; sandbox for workers | detected; prevented only inside a sandbox |
| Unauthorized operator | engage or lift the kill switch, replay a control message | pinned operator key, nonces | refused |
| Other tenant | read or inject evidence, reuse evidence across tenants | tenant binding in every signed body; store tokens | refused, 0 leaks |
| Untrusted content | become trusted memory, steer a strategy | source and status rules, instruction screen, held-out guard | quarantined |
| Tampering with evidence | edit a receipt, a row, an event | signatures, hash chains, Merkle root | detected |
| The agent approving itself | sign its own approval or promote its own strategy | approver must be a pinned human key / a different identity | refused |

33/33 lab scenarios refused; 34 tests pass. All testing is by the same project.

## Not defended, stated plainly

1. **Collusion of pinned keys.** If the authorizer, enforcement point and resource keys are held by one compromised
   party, a consistent false receipt can be produced. There is no independent witness.
2. **Root on the host.** Keys are files; there is no hardware root and no attestation.
3. **Same-user file access outside a sandbox.** Detected, not prevented.
4. **Network partition.** A partitioned reader fails closed, but kill propagation across hosts is not built.
5. **The remote resource** does not check the kill switch itself.
6. **Paraphrased prompt injection** passes the lexical screens in memory and in the MCP scanner (measured misses).
7. **Model-level attacks** (jailbreaks, hidden behaviours) are out of scope: governance bounds what a model may do,
   not what it thinks.
8. **The hosted production path** does not run this code.
9. **Verifier independence.** The offline verifier is written by the same team that produced the evidence.
