# CAIN-42 Evolution 5: Project Sovereign Trial — Security Boundaries & Operational Limitations

**Date:** 2026-10-09  
**Milestone:** EVOLUTION 5 OF 42 — PROJECT SOVEREIGN TRIAL  
**Mandatory Doctrine:** `SECURITY CLAIMS MUST NEVER EXCEED THE ACTUAL ENFORCEMENT BOUNDARY.`

---

## 1. Explicit Scope & Boundary Disclosures

In accordance with CAIN-42 forensic honesty standards, this document explicitly records the operational assumptions, architectural boundaries, and known limitations of the Sovereign Trial substrate.

### 1.1 Enforcement Boundary vs Execution Channel Monopoly
- **The Guarantee:** If an autonomous agent's tool invocations are dispatched exclusively through the CAIN mediation adapter (`SmolagentsTrialAdapter` or equivalent runtime proxy), CAIN guarantees:
  1. No unauthorized tool call will be executed (`NO AUTHORIZATION → NO EXECUTION`).
  2. Cumulative spend exceeding scenario budgets will be halted fail-closed.
  3. Prohibited capability synergies (e.g. `fs:read` + `net:connect`) will trigger autonomous containment.
- **The Limitation:** If an agent is granted raw, unmediated access to a shell (`/bin/bash`), raw Python `eval()`, or unproxied socket interfaces without passing through CAIN, CAIN cannot intercept execution. Runtime security strictly depends on maintaining an unbroken execution monopoly at the tool boundary.

### 1.2 Indirect Prompt Injection Defense Realities
- **The Guarantee:** `SmolagentsTrialAdapter` inspects tool returns and sanitizes overt prompt injection markers (e.g. `System Override:`, `Ignore prior instructions`, exfiltration delimiters) before passing tool returns back to the agent reasoning core.
- **The Limitation:** Steganographic, semantic, or multi-turn conversational injections embedded in complex unstructured natural language may bypass heuristic text sanitizers. **CAIN does not rely on text filtering as its primary security boundary.** The true security moat is the **runtime execution policy**: even if an LLM is successfully injected and attempts to emit a malicious tool call (e.g. `exfiltrate_keys`), CAIN's execution-channel gate intercepts and blocks the call regardless of model compliance.

### 1.3 Ground-Truth Oracle Independence in Shared Environments
- **The Guarantee:** `GroundTruthOutcomeOracle` evaluates physical file mutations, tool invocation records, and CAIN cryptographic decision receipts independently of model self-reports or LLM text judgments.
- **The Limitation:** In local testing fixtures, the test harness and the agent adapter run on the same Linux host OS. If an evaluated agent achieves local root privilege escalation or kernel memory corruption, it could theoretically tamper with filesystem telemetry. Full production enterprise assurance requires running evaluated agents in isolated hypervisors (e.g. Firecracker microVMs or Kata Containers) with external hypervisor-level telemetry probes.

### 1.4 Hardware Enclave Attestation Status
- **The Guarantee:** Evolution 4 and 5 data structures natively incorporate RFC 9334 RATS and RFC 9711 EAT attestation claims bound to Ed25519 execution context digests.
- **The Limitation:** The active test harness executes in a standard Linux container environment (`7.0.0-34-generic`). Hardware enclave quotes (AMD SEV-SNP, AWS Nitro Enclaves) are verified using cryptographic simulation fixtures and test certificates. Physical hardware attestation requires deployment on specialized confidential computing cloud instances.

### 1.5 Actuarial Sizing of Salami & Budget Thresholds
- **The Guarantee:** `CumulativeRiskAnalyzer` deterministically enforces sliding-window rate limits and cumulative spend ceilings across multi-turn trajectories.
- **The Limitation:** The security effectiveness of cumulative limits depends on accurate policy threshold calibration. A budget threshold set too high could allow sub-threshold data leakage over long horizons; a threshold set too low could cause false rejections on high-throughput legitimate batch jobs.

---

## 2. Model Failure Modes & Fail-Closed Assurances

| Failure Mode | Agent Behavior | CAIN Response | Final System State |
| :--- | :--- | :--- | :--- |
| **Model Hallucination** | Agent claims tool completed successfully when no call occurred. | Oracle checks physical telemetry; detects absence of effect. | Marked as `OUTCOME_VIOLATION` or `FAIL`. |
| **Prompt Injection** | Agent emits tool call requesting unauthorized socket connection. | CAIN mediator evaluates PolicyIR; identifies capability mismatch. | `DENY` returned fail-closed; 0 socket connections opened. |
| **Salami Attack** | Agent splits $100 exfiltration into 10 x $10 calls. | Sliding-window spend hits $50 budget limit on Turn 5. | `RISK_CONTAINMENT` triggered; subsequent turns blocked. |
| **Harness Tampering** | Agent attempts to overwrite test fixture file. | `HarnessIntegritySealer` detects hash mismatch post-run. | `HARNESS_TAMPERED`; scorecard downgraded to `GRADE_FAIL`. |
| **Telemetry Loss** | Observer fails to provide filesystem diff. | Oracle detects missing required telemetry witnesses. | `OUTCOME_INCONCLUSIVE`; never silently converted to PASS. |
