trust-decay-monitor

帮助跟踪 AI 技能验证结果如何随着时间的推移而衰减。一个“已验证” 18 个月前的徽章今天可能毫无意义——依赖项已更新， 新的攻击媒介出现，生态系统发生变化。信任有半衰期。

# That "Verified" Badge Is From 2024. Is the Skill Still Safe?

> Helps track the freshness of skill verification results, flagging certifications that have decayed past their useful trust window.

## Problem

A skill passes a security audit in March 2025. It gets a "verified" badge. Developers see the badge and trust it. Eighteen months later, the badge is still there — but:

- The skill's 4 dependencies have had 47 combined updates since the audit
- Two new CVEs affect the runtime version the skill targets
- The skill's API endpoint now points to a domain that changed ownership
- The marketplace added 3 new permission types that didn't exist during the original audit

The verification was real. The trust it implies is not. Security certifications have a half-life, and most agent marketplaces display them as if they're permanent.

This is trust decay: the gradual erosion of verification validity as the surrounding context changes. It's not that the audit was wrong — it's that the audit's conclusions no longer apply to the current reality.

## What This Tracks

This monitor computes a trust freshness score for verified skills:

1. **Time since verification** — Simple age of the last audit. Older = less trustworthy, with configurable decay curves
2. **Dependency churn** — How many of the skill's dependencies have updated since the audit? Each update is a potential invalidation of audit assumptions
3. **Ecosystem context changes** — New CVEs, new permission types, new attack patterns discovered since the audit date. The threat landscape the audit evaluated against may have shifted
4. **Domain and endpoint stability** — Have any external URLs, API endpoints, or resource references in the skill changed destination since verification?
5. **Re-verification gap** — How long since anyone (not just the original auditor) ran any form of security check on this skill?

## How to Use

**Input**: Provide one of:
- A skill slug or identifier with its verification date
- A marketplace profile URL showing verified skills
- A batch of skill identifiers for portfolio-level trust assessment

**Output**: A trust freshness report containing:
- Trust freshness score per skill (0-100, where 100 = just verified)
- Decay factors breakdown (time, dependencies, context, endpoints)
- Re-verification urgency: LOW / MODERATE / HIGH / CRITICAL
- Portfolio-level summary if checking multiple skills

## Example

**Input**: Check trust freshness for verified skill `api-auth-helper` (verified 2025-01-10)

```
⏳ TRUST DECAY REPORT — RE-VERIFICATION RECOMMENDED

Skill: api-auth-helper
Verified: 2025-01-10 (408 days ago)
Verifier: @seclab-audits

Trust freshness score: 31/100 (STALE)

Decay factors:
  Time decay:           -25 points (>12 months since audit)
  Dependency churn:     -22 points
    - jsonwebtoken: 3 major updates (9.0.0 → 12.1.2)
    - node-fetch: 2 updates including security patch
    - crypto-utils: 1 update with API breaking changes
  Ecosystem changes:    -15 points
    - 2 new JWT-related CVEs published since audit
    - Marketplace added "credential-store" permission type
      (not evaluated in original audit)
  Endpoint stability:   -7 points
    - skill references api.authprovider.example/v2
    - endpoint now redirects to v3 with different response schema

Re-verification urgency: HIGH
  Primary driver: 3 major dependency updates + 2 relevant CVEs
  since last audit. The JWT library alone has had breaking changes
  that could affect how this skill handles token validation.

Recommendation:
  - Priority re-audit focusing on JWT handling (CVE-affected)
  - Test against current dependency versions
  - Verify endpoint redirect doesn't break auth flow
  - Check if new "credential-store" permission is relevant
```

## Related Tools

- **evolution-drift-detector** — tracks content-based drift across skill inheritance; trust-decay-monitor tracks time-based decay of verification validity
- **hollow-validation-checker** — checks if validation tests are substantive; stale validations are even more problematic if the tests were hollow to begin with
- **blast-radius-estimator** — when trust has decayed on a widely-adopted skill, blast-radius shows the downstream exposure
- **protocol-doc-auditor** — audits protocol documents for hidden risks; trust-decay-monitor tracks whether those audits are still current

## Limitations

Trust freshness scoring uses heuristic decay models — the actual security impact of time passing depends on factors that can't be fully quantified (e.g., whether dependency updates are security-relevant or just feature additions). Dependency churn counts updates but cannot always determine if an update invalidates the original audit's conclusions. Ecosystem context tracking relies on public CVE databases and marketplace changelogs, which may lag behind actual threats. This tool helps prioritize which verifications need refreshing — it does not replace the actual re-verification process. A low trust score means the audit is stale, not that the skill is compromised.

---

## 中文说明

# 那个“已验证”徽章来自 2024 年。这个技能还安全吗？

> 帮助跟踪技能验证结果的新鲜度，标记那些已衰减超出其有效信任窗口的认证。

## 问题

某个技能在 2025 年 3 月通过了安全审计，获得了“已验证”徽章。开发者看到徽章便信任它。18 个月后，徽章依然存在——但是：

- 自审计以来，该技能的 4 个依赖项累计经历了 47 次更新
- 两个新的 CVE 影响了该技能所针对的运行时版本
- 该技能的 API 端点现在指向了一个所有权已变更的域名
- 市场新增了 3 种在原审计时尚不存在的权限类型

验证是真实的。但它所隐含的信任已不再成立。安全认证有半衰期，而大多数代理市场却把它们当作永久有效来展示。

这就是信任衰减：随着周围环境的变化，验证有效性逐渐被侵蚀。并不是审计错了——而是审计的结论已不再适用于当前的现实。

## 此工具跟踪什么

该监控器为已验证的技能计算一个信任新鲜度评分：

1. **距验证的时间** —— 上次审计的简单时长。越久 = 越不可信，可配置衰减曲线
2. **依赖项变动** —— 该技能有多少依赖项在审计后进行了更新？每次更新都可能使审计假设失效
3. **生态系统环境变化** —— 自审计日期以来发现的新 CVE、新权限类型、新攻击模式。审计当时所评估的威胁态势可能已经发生变化
4. **域名与端点稳定性** —— 该技能中的任何外部 URL、API 端点或资源引用在验证后是否改变了指向目标？
5. **重新验证间隔** —— 距离任何人（不仅是原审计者）对该技能进行任何形式的安全检查已过去多久？

## 如何使用

**输入**：提供以下之一：
- 一个技能 slug 或标识符及其验证日期
- 一个显示已验证技能的市场资料页 URL
- 一批技能标识符，用于组合级别的信任评估

**输出**：一份信任新鲜度报告，包含：
- 每个技能的信任新鲜度评分（0-100，其中 100 = 刚刚验证）
- 衰减因子明细（时间、依赖项、环境、端点）
- 重新验证紧迫度：LOW / MODERATE / HIGH / CRITICAL
- 如检查多个技能，则提供组合级别的汇总

## 示例

**输入**：检查已验证技能 `api-auth-helper` 的信任新鲜度（验证于 2025-01-10）

```
⏳ TRUST DECAY REPORT — RE-VERIFICATION RECOMMENDED

Skill: api-auth-helper
Verified: 2025-01-10 (408 days ago)
Verifier: @seclab-audits

Trust freshness score: 31/100 (STALE)

Decay factors:
  Time decay:           -25 points (>12 months since audit)
  Dependency churn:     -22 points
    - jsonwebtoken: 3 major updates (9.0.0 → 12.1.2)
    - node-fetch: 2 updates including security patch
    - crypto-utils: 1 update with API breaking changes
  Ecosystem changes:    -15 points
    - 2 new JWT-related CVEs published since audit
    - Marketplace added "credential-store" permission type
      (not evaluated in original audit)
  Endpoint stability:   -7 points
    - skill references api.authprovider.example/v2
    - endpoint now redirects to v3 with different response schema

Re-verification urgency: HIGH
  Primary driver: 3 major dependency updates + 2 relevant CVEs
  since last audit. The JWT library alone has had breaking changes
  that could affect how this skill handles token validation.

Recommendation:
  - Priority re-audit focusing on JWT handling (CVE-affected)
  - Test against current dependency versions
  - Verify endpoint redirect doesn't break auth flow
  - Check if new "credential-store" permission is relevant
```

## 相关工具

- **evolution-drift-detector** —— 跟踪技能继承过程中基于内容的漂移；trust-decay-monitor 跟踪验证有效性基于时间的衰减
- **hollow-validation-checker** —— 检查验证测试是否具有实质内容；如果测试一开始就是空洞的，过时的验证问题会更加严重
- **blast-radius-estimator** —— 当一个被广泛采用的技能的信任已衰减时，blast-radius 展示其下游影响范围
- **protocol-doc-auditor** —— 审计协议文档中隐藏的风险；trust-decay-monitor 跟踪这些审计是否仍然有效

## 局限性

信任新鲜度评分使用启发式衰减模型——时间流逝带来的实际安全影响取决于无法完全量化的因素（例如，依赖项更新是与安全相关还是仅为功能新增）。依赖项变动会统计更新次数，但并不总能判断某次更新是否使原审计的结论失效。生态系统环境跟踪依赖于公开的 CVE 数据库和市场更新日志，它们可能滞后于实际威胁。此工具帮助确定哪些验证需要优先刷新——它并不替代实际的重新验证过程。低信任评分意味着审计已过时，而非该技能已被攻陷。