AIDLC Security Agent

Enterprise-oriented Security Agent for AI-driven development lifecycle (AIDLC) workflows. It is designed to reduce security issues introduced by vibe coding, review pull requests before CI/CD deployment, ingest CVE data from application MCPs, and enforce human-in-the-loop approval for high-risk work.

What It Provides

Extensible MCP integration: register GitHub, Prisma, or any future MCP through MCPRegistry or JSON config.
PR review flow: combines local repository scanning with optional GitHub MCP findings.
CI/CD gate: fails deployment gates on high or critical findings until fixed or explicitly approved.
CVE backtracking: maps CVE records returned by MCPs to local dependency manifests and plans upgrades.
Controlled remediation: high-risk CVE changes require human approval before code or manifest modification.
LLM-assisted left shift: uses OpenAI-compatible or Anthropic-compatible APIs to turn findings into prioritized AIDLC security work.
Admin console: built-in web UI for command output streaming, task scheduling, agent assignment, skill install/configuration, and approvals.
Offline testability: static MCP connectors make policy and CI behavior reproducible without external services.

Architecture

security_agent/
  agent.py        High-level orchestration API
  approvals.py    Human-in-the-loop approval store and enforcement
  cli.py          CLI for review, gate, CVE planning, CVE apply, approvals
  cve.py          CVE parsing, dependency mapping, remediation planning/apply
  llm.py          OpenAI-compatible and Anthropic-compatible HTTP adapters
  left_shift.py   LLM-backed AIDLC left-shift analysis
  admin.py        Built-in web admin server with realtime task output
  mcp/            MCP connector abstraction and stdio/static adapters
  pipeline.py     CI/CD gate evaluation
  policy.py       Enterprise blocking and approval policy
  reviewer.py     PR/local/MCP review orchestration
  scanners.py     Local secret, unsafe code, and dependency scanning

Quick Start

python -m pip install -e .
security-agent init-config --path security-agent.json
security-agent review-pr --repo . --output security-report.json
security-agent gate --report security-report.json
security-agent admin --repo . --host 127.0.0.1 --port 8765

Exit codes:

0: review/gate passed or command succeeded.
1: review/gate found blocking security issues.
2: remediation requires human approval.
3: remediation was rejected.

MCP Connectors

Connectors implement a narrow call_tool(tool_name, arguments) interface. The included adapters are:

StdioMCPConnector: calls MCP servers over JSON-RPC stdio.
StaticMCPConnector: deterministic test/dry-run connector.

Example config:

{
  "approvals_path": ".security-agent/approvals.json",
  "policy": {
    "block_on": ["high", "critical"],
    "require_approval_for": ["high"],
    "fail_on_missing_fixed_version": true
  },
  "mcp_connectors": [
    {"name": "github", "type": "stdio", "command": ["github-mcp-server"]},
    {"name": "prisma", "type": "stdio", "command": ["prisma-mcp-server"]}
  ]
}

To add a new MCP later, add a config entry or implement MCPConnector for custom auth, transport, or enterprise sandboxing.

LLM Providers

The agent supports two API payload formats without requiring provider SDKs:

type: "openai": OpenAI-compatible /chat/completions APIs.
type: "anthropic": Anthropic-compatible /v1/messages APIs.

Example:

{
  "default_llm_provider": "openai-main",
  "llm_providers": [
    {
      "name": "openai-main",
      "type": "openai",
      "model": "openai-compatible-security-model",
      "base_url": "https://api.openai.com/v1",
      "api_key_env": "OPENAI_API_KEY"
    },
    {
      "name": "anthropic-main",
      "type": "anthropic",
      "model": "anthropic-compatible-security-model",
      "base_url": "https://api.anthropic.com",
      "api_key_env": "ANTHROPIC_API_KEY"
    }
  ]
}

Run left-shift analysis after a review:

security-agent --config examples/security-agent.json advise --report security-report.json --provider openai-main

The model is used for prioritization and remediation planning. It does not bypass the deterministic policy gate or human approval checks.

Human-In-The-Loop

High-risk actions are blocked at execution time by ApprovalStore.ensure_allowed. For example, applying a high or critical CVE upgrade first creates an approval request:

security-agent cve --repo . --payload examples/cve-payload.json --apply
security-agent approvals list
security-agent approvals approve <approval-id> --by alice@example.com --reason "Reviewed patch and test scope"
security-agent cve --repo . --payload examples/cve-payload.json --apply

This keeps high-risk code modifications, deployment unblocking, and risk acceptance under human control.

For PR findings that require explicit risk acceptance before deployment:

security-agent approvals request --report security-report.json --requester ci-security-gate
security-agent approvals approve <approval-id> --by security-lead@example.com --reason "Accepted for this deployment window"
security-agent gate --report security-report.json

GitHub PR And CI/CD Usage

The included workflow in .github/workflows/security-agent.yml installs the package, reviews the repository, writes security-report.json, and enforces the deployment gate before later pipeline stages. In production, provide a GitHub MCP connector with repository and PR context so MCP findings are merged with local scanner results.

Admin Console

Start the built-in management system:

security-agent --config examples/security-agent.json admin --repo . --host 127.0.0.1 --port 8765

Open http://127.0.0.1:8765.

The console provides:

Realtime command output using Server-Sent Events.
Task scheduling and execution for review, gate, CVE, and custom commands.
Agent records with role, LLM provider, MCP connector, and skill assignments.
Skill installation/configuration. Local skill folders or files are copied into .security-agent/skills; external/manual skills can be recorded as configuration-only entries.
Human approval list with approve/reject actions.

CVE Payload Contract

Third-party MCPs can return:

{
  "cves": [
    {
      "cve_id": "CVE-2026-0001",
      "package": "example-lib",
      "affected_versions": "<2.0.0",
      "fixed_version": "2.0.0",
      "severity": "critical",
      "description": "Example vulnerability"
    }
  ]
}

The agent maps package names to package.json, requirements*.txt, and pyproject.toml dependencies. If a fixed version is present, it plans an upgrade. If no fixed version exists, it creates a high-risk manual mitigation action.

Development

python -m unittest discover -s tests

AIDLC Security Agent 中文版

面向企业的 AIDLC（AI 驱动开发生命周期）安全代理。它用于降低 vibe coding 或 AI 生成代码带来的安全风险，在 CI/CD 部署前审查 pull request，从应用 MCP 中接收 CVE 数据，并对高风险操作执行人工审批。

提供能力

可扩展的 MCP 集成：可通过 MCPRegistry 或 JSON 配置注册 GitHub、Prisma 或后续任意 MCP。
PR 审查流程：结合本地仓库扫描和可选的 GitHub MCP 发现结果。
CI/CD 安全门禁：当存在 high 或 critical 级别问题时阻止部署，直到问题修复或明确批准。
CVE 回溯：把 MCP 返回的 CVE 记录映射到本地依赖清单，并生成升级计划。
受控修复：高风险 CVE 变更在修改代码或依赖清单前需要人工批准。
LLM 辅助左移：使用 OpenAI 兼容或 Anthropic 兼容 API，把安全发现转化为有优先级的 AIDLC 安全工作项。
管理控制台：内置 Web UI，支持命令输出流、任务调度、代理分配、技能安装/配置和审批处理。
离线可测试：静态 MCP 连接器让策略和 CI 行为可以在无外部服务环境中复现。

架构

security_agent/
  agent.py        高层编排 API
  approvals.py    人工审批存储和执行控制
  cli.py          用于审查、门禁、CVE 计划、CVE 应用和审批的 CLI
  cve.py          CVE 解析、依赖映射、修复计划和应用
  llm.py          OpenAI 兼容与 Anthropic 兼容 HTTP 适配器
  left_shift.py   基于 LLM 的 AIDLC 左移分析
  admin.py        内置 Web 管理服务，支持实时任务输出
  mcp/            MCP 连接器抽象，以及 stdio/static 适配器
  pipeline.py     CI/CD 门禁评估
  policy.py       企业阻断和审批策略
  reviewer.py     PR、本地和 MCP 审查编排
  scanners.py     本地密钥、危险代码模式和依赖扫描

快速开始

python -m pip install -e .
security-agent init-config --path security-agent.json
security-agent review-pr --repo . --output security-report.json
security-agent gate --report security-report.json
security-agent admin --repo . --host 127.0.0.1 --port 8765

退出码：

0：审查/门禁通过，或命令成功。
1：审查/门禁发现阻断性安全问题。
2：修复操作需要人工审批。
3：修复操作被拒绝。

MCP 连接器

连接器实现一个很窄的 call_tool(tool_name, arguments) 接口。内置适配器包括：

StdioMCPConnector：通过 JSON-RPC stdio 调用 MCP 服务。
StaticMCPConnector：用于测试和 dry-run 的确定性连接器。

示例配置：

{
  "approvals_path": ".security-agent/approvals.json",
  "policy": {
    "block_on": ["high", "critical"],
    "require_approval_for": ["high"],
    "fail_on_missing_fixed_version": true
  },
  "mcp_connectors": [
    {"name": "github", "type": "stdio", "command": ["github-mcp-server"]},
    {"name": "prisma", "type": "stdio", "command": ["prisma-mcp-server"]}
  ]
}

后续添加新的 MCP 时，可以增加配置项，也可以实现 MCPConnector 来处理自定义认证、传输或企业沙箱策略。

LLM 提供方

代理支持两种 API 请求格式，不需要安装提供方 SDK：

type: "openai"：OpenAI 兼容的 /chat/completions API。
type: "anthropic"：Anthropic 兼容的 /v1/messages API。

示例：

{
  "default_llm_provider": "openai-main",
  "llm_providers": [
    {
      "name": "openai-main",
      "type": "openai",
      "model": "openai-compatible-security-model",
      "base_url": "https://api.openai.com/v1",
      "api_key_env": "OPENAI_API_KEY"
    },
    {
      "name": "anthropic-main",
      "type": "anthropic",
      "model": "anthropic-compatible-security-model",
      "base_url": "https://api.anthropic.com",
      "api_key_env": "ANTHROPIC_API_KEY"
    }
  ]
}

审查完成后运行左移分析：

security-agent --config examples/security-agent.json advise --report security-report.json --provider openai-main

模型只用于优先级排序和修复规划建议。它不会绕过确定性的策略门禁或人工审批检查。

人工审批

高风险操作会在执行时被 ApprovalStore.ensure_allowed 阻断。例如，应用 high 或 critical 级别的 CVE 升级时会先创建审批请求：

security-agent cve --repo . --payload examples/cve-payload.json --apply
security-agent approvals list
security-agent approvals approve <approval-id> --by alice@example.com --reason "Reviewed patch and test scope"
security-agent cve --repo . --payload examples/cve-payload.json --apply

这让高风险代码修改、部署放行和风险接受始终处于人工控制之下。

如果 PR 发现项需要在部署前显式接受风险：

security-agent approvals request --report security-report.json --requester ci-security-gate
security-agent approvals approve <approval-id> --by security-lead@example.com --reason "Accepted for this deployment window"
security-agent gate --report security-report.json

GitHub PR 与 CI/CD 用法

.github/workflows/security-agent.yml 中的工作流会安装本包、审查仓库、写入 security-report.json，并在后续流水线阶段前执行部署门禁。生产环境中，应提供带有仓库和 PR 上下文的 GitHub MCP 连接器，使 MCP 发现项可以与本地扫描结果合并。

管理控制台

启动内置管理系统：

security-agent --config examples/security-agent.json admin --repo . --host 127.0.0.1 --port 8765

打开 http://127.0.0.1:8765。

控制台提供：

基于 Server-Sent Events 的实时命令输出。
审查、门禁、CVE 和自定义命令的任务调度与执行。
包含角色、LLM 提供方、MCP 连接器和技能分配的代理记录。
技能安装/配置。本地技能目录或文件会复制到 .security-agent/skills；外部或手工技能可作为仅配置项记录。
人工审批列表，以及批准/拒绝操作。

CVE Payload 合同

第三方 MCP 可以返回：

{
  "cves": [
    {
      "cve_id": "CVE-2026-0001",
      "package": "example-lib",
      "affected_versions": "<2.0.0",
      "fixed_version": "2.0.0",
      "severity": "critical",
      "description": "Example vulnerability"
    }
  ]
}

代理会把包名映射到 package.json、requirements*.txt 和 pyproject.toml 中的依赖。如果存在 fixed version，它会计划升级；如果没有 fixed version，它会创建高风险的人工缓解操作。

开发

python -m unittest discover -s tests

Name		Name	Last commit message	Last commit date
Latest commit History 2 Commits
.claude		.claude
.github/workflows		.github/workflows
examples		examples
security_agent		security_agent
tests		tests
.gitignore		.gitignore
README.md		README.md
pyproject.toml		pyproject.toml

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

AIDLC Security Agent

What It Provides

Architecture

Quick Start

MCP Connectors

LLM Providers

Human-In-The-Loop

GitHub PR And CI/CD Usage

Admin Console

CVE Payload Contract

Development

AIDLC Security Agent 中文版

提供能力

架构

快速开始

MCP 连接器

LLM 提供方

人工审批

GitHub PR 与 CI/CD 用法

管理控制台

CVE Payload 合同

开发

About

Uh oh!

Releases

Packages

Uh oh!

Contributors

Uh oh!

Languages

Folders and files

Latest commit

History

Repository files navigation

AIDLC Security Agent

What It Provides

Architecture

Quick Start

MCP Connectors

LLM Providers

Human-In-The-Loop

GitHub PR And CI/CD Usage

Admin Console

CVE Payload Contract

Development

AIDLC Security Agent 中文版

提供能力

架构

快速开始

MCP 连接器

LLM 提供方

人工审批

GitHub PR 与 CI/CD 用法

管理控制台

CVE Payload 合同

开发

About

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages