AI security and authorization framework for autonomous agents
Alpha Sentinel is a security-first, deterministic control plane designed to separate AI reasoning from execution authority.
Core principle
AI proposes. Deterministic controls decide. Cryptography binds. Isolation signs. Blockchain verifies.
The principle describes the target architecture. It is not a claim that every stage is operationally integrated or production-verified.
Autonomous AI agents can interact with consequential systems, including blockchain, DeFi, and financial automation. Model-generated decisions must not serve as the final authorization boundary.
Alpha Sentinel explores independently enforceable deterministic controls before consequential actions are authorized.
- Deterministic risk assessment and policy enforcement
- Transaction simulation and re-simulation gates in the authorization pipeline
- Transaction-bound approvals and timelock enforcement
- An isolated-signer interface with a fail-closed default implementation; operational key isolation is not yet demonstrated
- Explicit lifecycle state machines, persistence, and recovery tests
- Audit and signed-evidence verification components, with end-to-end verification still under evaluation
| Layer | Responsibility and current boundary |
|---|---|
| Policy Engine | Deterministic risk and authorization decisions |
| Simulation Boundary | Validate a transaction using an injected simulator before signing; real-chain guarantees require a trusted implementation |
| Approval Binding | Bind approval to exact proposal and transaction digests |
| Isolated Signing | Signer abstraction exists; default signer refuses to sign, and operational isolated signing is not demonstrated |
| Lifecycle State Machine | Explicit transitions and recovery/replay controls |
| Audit & Verification | Evidence checks and regression tests; live broadcast and independent receipt verification are not demonstrated end to end |
Alpha Sentinel is under active development. Core control-plane components, authorization boundaries, lifecycle handling, and regression tests are implemented. Production readiness and full end-to-end security guarantees remain validation objectives.
At signed default-branch commit b9f2d8d815a13362717e6d3eaf36cdb0d6650112, GitHub Actions recorded:
- 171/171 top-level Node tests passing
- 172/172 recursive Node tests passing (overlapping with the top-level suite; do not add the counts)
- 26/26 Python security-boundary tests passing
Evidence: Coverage v1 · security-boundary · ML Pipeline CI.
PR #50, merged as 7504edb09418766a5046a9789c879a01a6549d8d, adds five bounded Python authorization regression tests and a dedicated Python 3.11 CI workflow. The dedicated PR workflow passed 5/5 tests. These tests use a recording signer test double, not an isolated production signer.
Not yet demonstrated: operational isolated key custody, a fully integrated live broadcast-and-receipt verification path, or end-to-end authorization guarantees under adversarial production conditions.
core/— Control-plane lifecycle and daemon componentssecurity/andsecurity_boundary/— Security evidence and Python authorization boundarysimulation/— Simulation boundary componentsstate/— State and lifecycle persistence componentstests/— Automated regression and security tests.github/workflows/— CI pipelines and dedicated authorization fixture gate
- Deterministic Agent Authorization — Test enforceable policies independently of model behavior
- Adversarial Robustness — Evaluate prompt injection, malformed outputs, and attempts to bypass execution controls
- Reproducible Security Evidence — Publish evaluation fixtures, regression coverage, and independently reviewable results
git clone https://github.com/cryptofixyup/alpha-sentinel.git
cd alpha-sentinel
# Run the bounded authorization fixture (requires Python 3.11 and pytest)
python -m pip install pytest
python -m pytest -q tests/test_grant_authorization_fixture.py
# Run the existing Python security-boundary suite
python -m pytest -q tests/test_security_boundary.pyThese are test fixtures. They do not perform production signing or live blockchain broadcasts.
A standalone monitoring tool checks disk paths, follows appended latency/memory metrics in watch mode, and can POST generic JSON alerts to a configured HTTPS receiver. It is informational and does not authorize actions. Target-host activation and metric-producer configuration remain explicit operational steps.
No repository license is currently declared in GitHub metadata, and no root LICENSE file was found during this update. Do not assume the source is licensed for unrestricted reuse. A license must be selected and added by the repository owner.
Piotr Kozłowski
Founder, CryptoFixYup
Independent software architect focused on deterministic systems, cryptographic integrity, and evidence-based security for autonomous AI.