hermes-security-harness - Hermes-native defensive security harness for autho

项目说明

Hermes Security Harness

Hermes-nati security harness for authorized, source-backed web targets.

The harness combines deterministic scanners, source-aware reconnaissance, structured artifact contracts, and optional Hermes/LLM-assisted reasoning. It is built for local and staging assessments where the operator controls the target and can provide explicit scope.

Safety posture

Current capabilities

| Area | Commands / modules | Notes |
|---|---|---|
| Target | | web-target/v1 YAML/JSON configs and safety gates. |
| HTTP smoke | http-smoke | Deterministic GET-only reachability plus security-header checks. |
| Source/static review | static-scan | Source in, threat model, structured findings, optional Hermes runner review. |
| Reconnaissance | recon | Forms, OpenAPI/Swagger, sitemap, JS bundle, hidden endpoints, filesystem/source-dir route discovery, staged unauth/auth support in Python API. |
| Injection testing | injection-scan | XSS, SQLi, SSRF payload testing with smoke/recon/auth surface support in the full scan flow. |
| Auth/session testing | auth-scan | Login flow, cookie checks, bypass tests, rate checks. |
| WSTG modules | csrf, http-verb, idor, jwt, stored-xss | OWASP WSTG-aligned focused scanners. |
| TLS | security_harness.tls_scan Python module | TLS/SSL configuration testing module. CLI wiring is not yet exposed. |
| Dependencies | dependency-audit | Lockfile parsing and C. |
| Rate limiting | rate-limit | Burst checks against configured endpoints. |
| PoC replay | replay-poc | Structured HTTP PoC replay with sandbox gates. |
| Chain correlation | chain | Deterministic chain correlation across findings. |
| Reporting | report | Structured report generation from findings JSON files. |
| Full scan | scan | Orchestrates smoke → static → auth → recon → injection → chain → LLM chain reasoning/report. |
| Async jobs | job-start, job-status, job-report, job-worker | JSON job registry/worker for gateway-safe polling. |

Install and test

If security-harness is on your PATH after editable install, the examples below can use security-harness directly. Otherwise use .venv/bin/python -m security_harness.cli.

Target config

Create a web-target/v1 config for an authorized local/staging target:

Quick smoke

http-smoke ne wildcard includePaths; add concrete paths such as /login or /api/health to the target config for broader reachability checks.

Full scan pipeline

The scan command is the preferred path when you want scanner outputs to feed later stages:

Current high-level flow:

1. HTTP smoke
2. Static/source scan
3. Auth scan
4. Recon
5. Injection scan using upstream smoke/recon/auth context where a
6. Deterministic chain correlation
7. LLM-assisted recursi reasoning/report artifacts where configured

This is the safest default because it avoids the common false-negati where scanners are run independently and ne discovered routes, forms, cookies, or smoke paths from earlier stages.

Architecture and data flow

The harness is organized around a small set of shared contracts and scanner modules:

1. Target config and safety — web_target.py loads web-target/v1 configs, allowed hosts, rejects unsafe en, and provides the source directory and scope used by scanners.
2. CLI dispatch — cli.py exposes the operator commands and adapts CLI arguments into scanner calls.
3. Full scan orchestration — scan_handler.py coordinates the scanner data flow for security-harness scan.
4. Scanner execution — indi scanner modules write JSON artifacts and return typed summaries where a.
5. Artifact consumers — chains.py, chain_reasoning.py, report.py, and the async jobs.py worker consume scanner findings and publish follow-on artifacts.

The most important runtime data flow is:

Focused WSTG scanners (csrf, http-verb, idor, jwt, stored-xss) use concrete endpoint paths from --endpoints when provided, otherwise they default to scope.includePaths from the target config. Detector names under detectors.enabled are not endpoint paths.

Focused scanner examples

PoC replay

Read-only PoCs can be replayed directly:

Mutation-capable PoCs require sandbox flags, concrete reset/seed lifecycle commands, an ephemeral home, a base-origin egress allowlist, and no credential mounts.

Chain correlation and reports

Reports include risk summaries, e, OWASP/MITRE-style mappings, C scoring, and remediation guidance where a.

LLM-assisted chaining

There are two chain-analysis paths:

1. security-harness chain is deterministic. It auto-tags findings and applies table-dri rules from security_harness/chains.py. This command does not call an LLM.
2. The full security-harness scan ... flow can run recursive LLM chain reasoning after deterministic chain correlation. The implementation lives in security_harness/chain_reasoning.py and is invoked from security_harness/scan_handler.py.

Current behavior in the full scan path:

Example full-scan in:

If no llm_chains.json appears, check the scan output. The current implementation skips LLM chain reasoning when deterministic chain correlation returns zero chains, even if indi findings exist.

Python API highlights

Prefer the scan CLI for normal use; use the Python API when composing custom workflows or tests.

Project structure

Artifact and git hygiene

Known remaining work

These are tracked as GitHub issues when identified during maintenance:

License

Apache-2.0

项目信息

开发语言：Python

Star 数：4

技术要点

该项目属于防护思路方向的开源资源，代码结构清晰，适合安全研究人员学习参考。项目围绕Hermes-native defensive security harness for authorized web targets展开，提供了可运行的实现，对理解相关安全机制、工具原理与攻防技术具有较高的研究价值。

从技术实现角度看，此类项目的核心在于对底层原理的深入理解与工程化落地。开发者在研究过程中可以结合官方文档与社区讨论，梳理项目的初始化流程、核心模块职责与数据流转，逐步掌握其中的技术细节与设计思路。

使用与学习建议

1. 下载源码压缩包，解压到本地环境。

2. 根据项目类型准备对应的运行环境与依赖。

3. 按照 README 中的步骤完成配置与部署。

4. 在隔离的测试环境中运行验证，避免影响生产系统。

免责声明

本资源来源于开源社区，仅供安全研究、技术学习与合法授权范围内的攻防演练使用，请勿用于任何非法或侵害他人权益的用途。使用本项目产生的一切后果由使用者自行承担。