Qoder Security：把安全审查内置进 AI 编程 Session，三层防护今日发布

Qoder Security: Three-Layer In-Session Code Security Built Natively into AI Coding

Tech-News

Jul 23, 2026

更新于 Jul 23, 2026

🇨🇳 中文

发布时间：2026-07-23
官网：qoder.ai/security
适用：Qoder Desktop + Qoder CLI
Qoder 用户规模：500万+ 注册开发者

一、问题：AI 写代码的速度已经超过安全审查的速度

2026年7月，一项针对近9000个C++程序的学术研究给出了一个不舒服的数字：AI 生成的代码触发已确认的运行时违规，约是人类写的代码的两倍——即使在控制了代码长度和测试通过率之后。

与此同时，主流的安全工具仍然跑在 CI 管道里，在代码提交之后才开始扫描。这意味着：当漏洞被发现的时候，开发者早已切换到下一个任务，上下文全没了，修复成本最高。

Qoder 的产品负责人 Ding Yu 对这个问题的表述很直白：

“We believe security cannot depend on a model’s good behavior. It has to be built into the product architecture.”

今天发布的 Qoder Security，是他们给出的答案。

二、核心思路：安全左移到代码出生的那一刻

Qoder Security 的设计目标是把安全审查从「提交后扫描」变成「编写时防护」——不是加一个外部插件，是原生内置进编程 session。

双 Agent 架构是关键设计决策：写代码的 Agent 和审查安全的 Agent 完全分离，避免「自审自查」问题。审查端进一步拆分为扫描 Agent 和验证 Agent 两个协作角色——发现问题之前先确认可达性，不是发现了就报。

三、三层防护：每一层做不同的事

Qoder Security 把安全工作分给三层，按深度递进：

L1 静态检查（Static Check）— 免费

时机：代码生成的瞬间
方式：模式匹配，规则驱动
做什么：危险函数调用出现的那一刻立即拦截并自动修复
特点：无延迟，无模型成本，对开发流程零打扰

硬编码密钥、已知危险 API 调用这类「确定性风险」在这一层处理掉。

L2 轻量扫描（Lightweight Scan）— 消耗 credits

时机：一个编码任务完成后，以建议形式出现
方式：LLM 语义理解 + 污点传播分析
做什么：理解代码在做什么，检测 SQL 注入、远程命令执行、敏感数据暴露等逻辑漏洞，先验证「可达性」再报告
特点：不打断编码流；只在自然断点出现，开发者决定是否触发

这一层是传统静态分析做不到的：规则无法理解逻辑，但 LLM 可以追踪数据从哪里来、到哪里去。

L3 深度扫描（Deep Scan）— 消耗 credits

时机：push 之前
方式：跨文件、跨函数的完整数据流追踪
做什么：从污点源追踪到危险 sink，发现需要跨越多个文件才能看见的链式漏洞
特点：单文件静态分析永远看不见的漏洞，在这一层才能浮现

四、闭环：发现 → 验证 → 修复 → 再验证

Qoder Security 不只是报告漏洞。流程是：

检测 → 验证可达性 → 生成修复建议 → Qoder 主 Agent 执行修复 → 下一轮扫描再验证

修复落地后，下一次扫描会自动复核。循环在 session 内部闭合，不留安全债。

五、数据

指标	数值（官方数据）
漏洞检测率提升	~60%（vs 传统工具）
误报率降低	~80%（vs 传统工具）
修复时间	数天 → 数分钟
内测发现	600+ 安全问题（生产级开源项目 + AI 基础设施组件）
Qoder 内部效果	代码审查中安全相关评论减少 35-45%

六、怎么开启

Qoder Desktop：Settings → Security → 开启

Qoder CLI：

/security-settings   # 配置安全设置
/security-scan       # 随时触发扫描
# 或直接用自然语言："scan this file for security issues"

Static Check 永久免费；Lightweight Scan 和 Deep Scan 消耗 credits。

七、判断

Qoder Security 做的事情有一个明确的名字：安全左移（shift left）。这不是新概念，但 AI 编程工具让这件事的时机提前到了极致——从「CI 管道」前移到「代码正在被写出来的那一秒」。

三层架构的设计是合理的：不是用同一个重型模型处理所有情况，而是让每一层做它最擅长的事——静态规则处理确定性风险，LLM 处理语义推理，深度扫描处理跨文件依赖。资源消耗也因此可控。

双 Agent 分离（写代码 vs 审查代码）是必要的工程决策。让同一个 Agent 审查自己写的代码，在理论上存在确认偏差。分离之后，审查 Agent 是独立的对抗角色。

80% 误报率降低这个数字值得关注。安全工具被开发者忽视的最大原因通常不是「漏报」，而是「报太多假警报」——每天刷几十条 false positive，开发者会养成无视报警的习惯，真正的漏洞也被淹没。把误报降下来，才能让安全审查真正融入工作流。

数据来源：Qoder 官网 qoder.ai/security，AccessNewswire 新闻稿，2026-07-23。

🇬🇧 English

Release date: 2026-07-23
Website: qoder.ai/security
Applies to: Qoder Desktop + Qoder CLI
Qoder user base: 5 million+ registered developers

1. The Problem: AI Writes Code Faster Than Security Review Can Keep Up

In July 2026, an academic study of nearly 9,000 C++ programs produced an uncomfortable statistic: AI-generated code triggers confirmed runtime violations at roughly twice the rate of human-written code — even after controlling for code length and test pass rates.

Meanwhile, mainstream security tools still run inside CI pipelines, only scanning after code is committed. This means: by the time a vulnerability is found, the developer has long since moved on to the next task — all context is gone, and the cost of remediation is at its highest.

Qoder’s product lead Ding Yu put it plainly:

“We believe security cannot depend on a model’s good behavior. It has to be built into the product architecture.”

Qoder Security, released today, is their answer.

2. Core Approach: Shift Security Left to the Moment Code Is Born

Qoder Security’s design goal is to transform security review from “post-commit scanning” into “in-session protection” — not as an external plugin, but natively built into the coding session itself.

The dual-Agent architecture is the key design decision: the code-writing Agent and the security-review Agent are completely separated, avoiding the “self-review” problem. The review side is further split into a scanning Agent and a validation Agent working in tandem — reachability is confirmed before a finding is reported, not reported on sight.

3. Three Layers of Defense: Each Layer Does Something Different

Qoder Security divides security work across three layers, ordered by increasing depth:

L1 Static Check — Free

Timing: The instant code is generated
Method: Pattern matching, rule-driven
What it does: Intercepts and auto-fixes dangerous function calls the moment they appear
Characteristics: Zero latency, no model cost, zero disruption to the development flow

“Deterministic risks” like hardcoded secrets and known dangerous API calls are handled at this layer.

L2 Lightweight Scan — Consumes Credits

Timing: After a coding task completes, appearing as suggestions
Method: LLM semantic understanding + taint propagation analysis
What it does: Understands what the code is doing, detects logic vulnerabilities such as SQL injection, remote command execution, and sensitive data exposure — verifies “reachability” before reporting
Characteristics: Does not interrupt the coding flow; only appears at natural breakpoints, and the developer decides whether to trigger it

This layer does what traditional static analysis cannot: rules cannot understand logic, but LLMs can track where data comes from and where it goes.

L3 Deep Scan — Consumes Credits

Timing: Before pushing
Method: Complete data-flow tracking across files and functions
What it does: Tracks from taint sources to dangerous sinks, uncovering chained vulnerabilities that are only visible by spanning multiple files
Characteristics: Vulnerabilities that single-file static analysis can never see emerge only at this layer

4. Closed Loop: Detect → Validate → Fix → Re-validate

Qoder Security does not merely report vulnerabilities. The workflow is:

检测 → 验证可达性 → 生成修复建议 → Qoder 主 Agent 执行修复 → 下一轮扫描再验证

Once a fix is applied, the next scan automatically re-verifies it. The loop closes within the session, leaving no security debt.

5. Data

Metric	Value (Official Data)
Vulnerability detection rate improvement	~60% (vs. traditional tools)
False positive rate reduction	~80% (vs. traditional tools)
Remediation time	Days → Minutes
Beta findings	600+ security issues (production open-source projects + AI infrastructure components)
Internal Qoder effect	Security-related comments in code review reduced by 35–45%

6. How to Enable

Qoder Desktop: Settings → Security → Enable

Qoder CLI:

/security-settings   # 配置安全设置
/security-scan       # 随时触发扫描
# 或直接用自然语言："scan this file for security issues"

Static Check is permanently free; Lightweight Scan and Deep Scan consume credits.

7. Assessment

What Qoder Security does has a clear name: shift left. This is not a new concept, but AI coding tools push the timing to its logical extreme — moving from “CI pipeline” all the way to “the very second code is being written.”

The three-layer architecture is sound: instead of applying the same heavyweight model to every situation, each layer does what it does best — static rules handle deterministic risks, LLMs handle semantic reasoning, deep scan handles cross-file dependencies. Resource consumption is therefore kept under control.

The dual-Agent separation (writing code vs. reviewing code) is a necessary engineering decision. Having the same Agent review code it wrote introduces theoretical confirmation bias. With separation, the review Agent is an independent adversarial role.

The 80% false positive reduction figure deserves attention. The main reason developers ignore security tools is usually not “missed detections” but “too many false alarms” — when developers wade through dozens of false positives every day, they develop a habit of ignoring alerts entirely, and real vulnerabilities get buried. Reducing false positives is what allows security review to genuinely integrate into the workflow.

Data source: Qoder official site qoder.ai/security, AccessNewswire press release, 2026-07-23.

💬 评论与讨论

使用 GitHub 账号登录后发表评论