thehardwarestore.ai

This repository is Microsoft's Agent Governance Toolkit, a production-grade multi-language framework (Python, TypeScript, Rust, Go, .NET) for enforcing policy, identity, trust, audit, and sandboxing on autonomous AI agents. In production, it intercepts agent tool calls, messages, and delegations before they reach external services, evaluating them against declarative policies (OPA, Cedar, or custom backends) with deterministic outcomes (allow/deny/escalate/transform). The system provides a full governance plane including a Rust-based policy engine (Agent Control Specification), a service mesh (AgentMesh) with API gateway, trust engine, policy server, and audit collector, a hypervisor for execution rings and sandboxing (Docker, Hyperlight, Nono), compliance verification (OWASP, EU AI Act), SLO monitoring, and a plugin marketplace. Integrations exist for LangChain, CrewAI, AutoGen, OpenAI Agents, Semantic Kernel, MCP, and many other frameworks. The toolkit is deployed as CLI tools (`agt`, `agentmesh`, `agentos`), Docker containers with sidecar patterns, Kubernetes charts, and VS Code/Chrome extensions. The architecture is polyglot with a shared policy engine core written in Rust; the Python SDK is the richest, with the .NET and Go SDKs covering governance middleware and lifecycle management. The codebase is heavily tested with 992 conformance tests across 10 formal specs, fuzzing targets, and CI gates enforcing deterministic policy outcomes, fail-closed behavior, and no-stub registration.

garak is a CLI tool and Python library for automated red-teaming and vulnerability scanning of large language models (LLMs). It systematically probes models with adversarial inputs (jailbreaks, prompt injections, encoding attacks, etc.) and uses detectors to classify model responses for safety failures. It supports multiple model backends (OpenAI, Anthropic, HuggingFace, etc.) and generates structured reports with severity scoring. The system follows a modular plugin architecture with base classes for probes, detectors, generators, harnesses, evaluators, and buffs. Plugins are dynamically loaded and configured via a hierarchical config system. The default harness runs probes with their associated detectors, while a cross-product harness allows all probes against all detectors. Results are logged as JSONL and can be visualized via a React-based web UI (garak-report) that displays DEFCON-level severity, calibration Z-scores, and technique-intent matrices. While primarily an application, garak incorporates significant research elements: it implements state-of-the-art attack techniques (GCG, AutoDAN, BEAST, TAP) and includes curated datasets of adversarial prompts and calibration data. The codebase serves both as a practical security tool and as a research platform for LLM vulnerability assessment.

enject is a security-focused CLI tool that prevents AI agents and other processes from reading secrets out of `.env` files. It stores secrets in an encrypted local store (AES-256-GCM with Argon2id key derivation) and injects them into subprocess environments at runtime. The `.env` file on disk contains only `en://` symbolic references, never plaintext values. The tool provides commands to initialize a store, add/update/list/delete secrets, import a plaintext `.env` file (encrypting all values and rewriting it as a template), rotate the master password, and run a subprocess with resolved environment variables. The architecture is a single Rust binary with a clean `Store` trait abstraction separating the encryption backend from command logic. The codebase demonstrates strong security practices: secrets never touch disk as plaintext, nonces are freshly generated on every write, key material is zeroized immediately after use, and unresolved references cause hard errors. It includes comprehensive unit tests and a CI pipeline with linting, formatting, and security auditing.

CyberClaw is an enterprise-grade transparent AI agent architecture that implements a dual-watermark memory system, full behavior auditing, zero-trust execution via two-phase skill invocation, and a heartbeat task engine for scheduled operations. The system provides an interactive conversational interface where users can query time, perform calculations, manage files, execute shell commands within a secure sandbox, and utilize pluggable skills. All agent actions are transparently logged to JSONL files for monitoring and compliance. The architecture features clean separation of concerns with distinct modules for agent logic, skill loading, context management, logging, heartbeat tasks, and tool execution.

This repository defines the Conversational Finance Governance Framework, an open specification (v0.4) for embedding runtime authority controls into conversational AI agents operating in fintech contexts. It includes a reference implementation (Lite sidecar) that demonstrates the control plane: classifying utterances along Financial Commitment Risk (F0-F3) and Regulatory Sensitivity (R0-R3) axes, binding each turn to an Action Object (inform/advise/commit/transact/represent), applying a decision table (FREE/AUTH/ESCALATE), enforcing classification assurance, issuing scoped credential records, and producing a hash-chained audit trail. The sidecar is a CLI tool that takes a product-line mapping YAML (payments, lending, insurance, collections) and an utterance, then outputs the governance decision as JSON. It does not call an LLM or execute real transactions; it is a control-plane demo. The repository also contains the full specification documents (v0.3 and v0.4), reference mapping files, and a test suite with gold-standard classification fixtures. The architecture is a monolithic CLI with a clean separation between classification, decision logic, and audit logging. The code is well-structured, uses dataclasses, and includes a reusable hash-chained audit log component. The last commit date (2026-09-09) appears to be in the future relative to the current date, which is anomalous but taken at face value indicates active maintenance.

DeepTeam is a production-oriented Python library and CLI for red-teaming LLM applications. It does not merely document vulnerabilities; it executes automated adversarial campaigns against a user-supplied model callback, generating attack prompts through LLM-driven templates (single-turn, multi-turn, tree/crescendo jailbreaks), scoring target responses with LLM-as-judge metrics, and aggregating outcomes into risk assessments with CVSS scoring. It also ships a code scanner for static AI-security review, a trace scanner for agent executions, and deployable input/output guardrails. Main entry points are the red_team() Python API and the deepteam CLI (login, run, scan), with YAML configuration driving model selection, vulnerability types, and attack budgets. Architecturally, the repo is a modular Python package built on deepeval: vulnerabilities inherit from BaseVulnerability, all metrics inherit from BaseRedTeamingMetric, attacks implement configurable templates and schemas, and frameworks (OWASP, MITRE ATLAS, NIST, EU AI Act, BeaverTails, Aegis) map risks to pre-built categories. Results flow through RTTestCase objects and can be saved locally as JSON or pushed to the Confident AI platform. The library is dual sync/async and includes progress reporting, cost accumulation, and retry logic. Although it embeds research-style synthetic data generation and dataset-driven framework integrations, the dominant character is an application library/CLI rather than a data research project. Strengths include broad vulnerability and attack coverage, clear base-class contracts, a substantial test suite, and active CI. Weaknesses include heavy copy-paste duplication across 30+ vulnerability, metric, and template modules, tight coupling to deepeval internals, reliance on hardcoded model defaults such as gpt-3.5-turbo-0125, and a known logic bug where is_successful in the metric base class compares instead of assigns, always returning False unless an error is set. Telemetry code also contains hardcoded service keys. Overall, it is a functional, well-documented framework with real reuse value, but it would benefit from deduplication and dependency hardening.

This is a cross-platform sandbox runtime library developed by Anthropic for securely executing untrusted commands within restricted environments. It is designed to be embedded by AI agent runtimes (such as Claude Code) to enforce network and filesystem access controls, perform TLS-terminating HTTP/SOCKS proxying, and mask credentials like API keys and AWS SigV4 secrets. The system operates as a filtering forward proxy with an optional external 'decider' process for per-request allow/deny/body-inspection verdicts, and it supports Linux (via bubblewrap + seccomp), macOS (via Seatbelt profiles), and Windows (via WFP and a Rust helper). The repository is primarily an application library with a CLI tool (`srt`), not a data or research project. It demonstrates high engineering quality with extensive test coverage, careful security invariants, and thoughtful handling of platform edge cases. However, its complexity is extreme—spanning three OS backends, custom proxy servers, TLS MITM, credential injection, and violation monitoring—and it carries significant drift risks from external binary dependencies, runtime-specific workarounds, and a large test surface with conditional skips.