SentinelReview

SentinelReview is a privacy-focused security review foundation for NEAR Rust smart contracts. It combines bounded source ingestion, deterministic checks, validated source evidence, and a NEAR AI inference adapter designed to keep AI-generated hypotheses separate from reproducible findings.

Project status (11 November 2026): The repository has a working security-analysis foundation and an implemented NEAR AI adapter. The project is still under active development; it is not yet a complete end-to-end auditing product, and no claim of independent NEAR approval or comprehensive contract safety is made.

Progress snapshot

Implemented foundations

• TypeScript project structure, versioned Zod contracts, and Vitest test suite.
• Bounded Rust source loader with path restrictions, file/byte limits, strict UTF-8 handling, and deterministic line accounting.
• Immutable source snapshots and canonical manifest/source hashing.
• Deterministic evidence validation for source identity, file paths, hashes, line ranges, and excerpts.
• Evidence integration with findings, including explicit unresolved/incomplete states.
• Initial deterministic NEAR Rust callback-result access-control rule. The rule is intentionally conservative: uncertain syntax or possible manual guards produce limitations/inconclusive results rather than a clean bill of health.
• NEAR AI adapter with strict candidate-response validation, source-grounded evidence handling, and separation of AI hypotheses from deterministic findings.
• SDK-backed attestation/response-verification handling and a controlled synthetic canary implementation with focused tests.
• Discovery diagnostics for inspecting NEAR AI gateway/model attestation and deployment measurements.

The repository includes focused tests for the implemented modules. Historical test runs were performed during earlier milestones; rerun the full suite and typecheck against the current checkout before treating the latest combined state as verified.

Current work in progress

• Compose deployment-image inventory parsing in the NEAR AI discovery report. Local work has started, including bounded input/service/image limits, YAML parsing, sanitized image-reference output, and malformed/unsupported input tests. This work is not yet present in the current main branch and must not be treated as completed or verified.
• Review the complete image-reference parser and its edge cases before merging that work.
• Re-run npm run typecheck and npm test after the current changes are finished.

Remaining roadmap

1. Deterministic security analysis

• Expand the NEAR Rust rule set beyond the initial callback-result rule.
• Record rule/tool versions and reproducible analysis metadata.
• Add broader safe-control and seeded-vulnerability fixtures.

2. Private AI analysis and provenance

• Complete a controlled provider canary using real, operator-supplied deployment measurements.
• Document exactly what provider attestation and response-signature checks establish—and what they do not.
• Keep independent NEAR approval distinct from operator-supplied measurement matching.

3. Reports and end-to-end workflow

• Define report generation, coverage accounting, and explicit limitations.
• Separate response verification, evidence validation, and report-integrity verification.
• Add end-to-end tests covering source intake through final report.

4. Security, reliability, and demo

• Expand tests for malformed input, prompt injection, timeouts, and provider failures.
• Enforce and test privacy-safe logging, access controls, and retention boundaries.
• Verify unavailable, failed, inconclusive, and not-run states can never appear as successful checks.
• Build a reproducible demo and document setup and operational requirements.

Current technical outline

• Contracts: src/contracts/v1.ts
• Source ingestion and hashing: src/source/load-source.ts, src/source/manifest-hash.ts
• Evidence validation and integration: src/source/validate-evidence.ts, src/source/integrate-evidence.ts
• Deterministic NEAR Rust rule: src/rules/near-rust-callback-private.ts
• NEAR AI adapter: src/ai/near-ai-adapter.ts
• Controlled canary: src/ai/near-ai-controlled-canary.ts
• Attestation discovery: scripts/near-ai-attestation-discovery.ts

Development commands

Use Node.js >=24.22.1 as declared in package.json.

npm install
npm run typecheck
npm test

Provider credentials belong only in a local/server environment. Never commit real API keys or sensitive deployment measurements.

Verification principles

• AI proposes; deterministic checks and validation decide what is presented as reproducible evidence.
• AI findings are hypotheses for human review; they do not establish exploitability.
• A missing finding does not mean a contract is safe or fully audited.
• A verified provider response signature does not prove analysis correctness or completeness.
• Operator-supplied deployment measurements are not independent NEAR approval.
• Evidence validation, provider attestation, response verification, and report integrity are separate properties.
• Failed, unavailable, inconclusive, and not-run checks must never be represented as passed.

Scope and limitations

SentinelReview is an evolving foundation, not a production-ready audit certification service. The current deterministic rule set is narrow, the full report/workflow layer remains incomplete, and the project does not claim exhaustive vulnerability detection, independent deployment approval, or a guarantee that reviewed contracts are safe.