Back to skill

Security audit

论衡 — 严肃长文流水线

Security checks for vulnerabilities and agentic risk

Overview

This is a disclosed multi-agent writing and research workflow with broad but purpose-aligned file, web, and session orchestration behavior.

Install only if you want a large, multi-agent writing workflow that will create many files, may spawn worker sessions, and may send search queries or URLs to external services after Phase 0 choices. Review the Phase 0 consent prompts carefully, especially external search, academic metadata, Firecrawl, operational telemetry, and all-auto mode.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (29)

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The document explicitly states that M-Integrity-2 at T7.5 must not read anything under final/, because those are T8-stage artifacts and should not exist yet. However, the pseudocode still reads final/交付说明.md and final/定稿.md, creating a contradictory trust boundary that can cause the gate to consume nonexistent, stale, or future-stage outputs and make invalid pass/fail decisions. In this pipeline context, that is dangerous because gating logic controls audit progression and can silently bypass stage isolation or fail closed on the wrong artifacts.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The file defines a security-relevant boundary ('禁读 final/') and then violates it in the adjacent pseudocode, which means reviewers or downstream agents cannot rely on the declared boundary. This kind of spec/implementation drift is hazardous in an agent pipeline because it can be exploited by causing the controller to read improper-stage artifacts, undermining deterministic orchestration and weakening audit integrity.

Content

No source excerpt is available for this finding.

Hidden Instructions

High
Category
Prompt Injection
Confidence
70% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · references/_shared/真源/pipeline-overview.md (reported line 59)May include surrounding context.

md
## 修订回环仲裁规则

<!-- generated: rounds-contract (do not edit by hand) -->
audit_revision.max_rounds: 2
t1b_targeted_review.max_rounds: 2
g14_style_gate.recheck_max_rounds: 2

Hidden Instructions

High
Category
Prompt Injection
Confidence
70% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · references/agents/00-主控-扩展职责.md (reported line 203)May include surrounding context.

md
| T2.5 完整性门 | M-Integrity-1 + M-Form-6 + M-Exist-3 | 数据卡完整 + 信任级别齐全 | status.md「M 门执行记录」 |
| T7.5 完整性门 | M-Integrity-2 + M-Form-8 + M-Exist-2 | 审计完整 + 三角验证覆盖 + 证据包齐全 | status.md「M 门执行记录」 |
| v1→v2 修订后 | **章节级 M 门**(按 `## ` 拆章节识别变更点)| 修订说明含「章节级变更检测」段 | 修订说明头部 |
| T8 终检 | M-Form 8 项 + M-Exist 3 项 = 11 项全过 | 四档判定 = 「通过」 | `final/M-Gate-Report-v2.2.12.json` | <!-- 注:`v2.2.12` 是 M-Gate 算法自有版本号(非 SKILL 版本),保持不变 -->

T8 终检前**分片必读** [`_shared/真源/M-Gate-核心.md`](../_shared/真源/M-Gate-核心.md) 的 **M-Form / M-Exist / M-Integrity 伪代码段**(附录部分按需 → [`_shared/真源/M-Gate-Algorithm-appendix.md`](../_shared/真源/M-Gate-Algorithm-appendix.md)),按伪代码执行,产出 `final/M-Gate-Report-v2.2.12.json`(标准 JSON 格式)。

Hidden Instructions

High
Category
Prompt Injection
Confidence
70% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · references/agents/00-主控-扩展职责.md (reported line 203)May include surrounding context.

md
| T2.5 完整性门 | M-Integrity-1 + M-Form-6 + M-Exist-3 | 数据卡完整 + 信任级别齐全 | status.md「M 门执行记录」 |
| T7.5 完整性门 | M-Integrity-2 + M-Form-8 + M-Exist-2 | 审计完整 + 三角验证覆盖 + 证据包齐全 | status.md「M 门执行记录」 |
| v1→v2 修订后 | **章节级 M 门**(按 `## ` 拆章节识别变更点)| 修订说明含「章节级变更检测」段 | 修订说明头部 |
| T8 终检 | M-Form 8 项 + M-Exist 3 项 = 11 项全过 | 四档判定 = 「通过」 | `final/M-Gate-Report-v2.2.12.json` | <!-- 注:`v2.2.12` 是 M-Gate 算法自有版本号(非 SKILL 版本),保持不变 -->

T8 终检前**分片必读** [`_shared/真源/M-Gate-核心.md`](../_shared/真源/M-Gate-核心.md) 的 **M-Form / M-Exist / M-Integrity 伪代码段**(附录部分按需 → [`_shared/真源/M-Gate-Algorithm-appendix.md`](../_shared/真源/M-Gate-Algorithm-appendix.md)),按伪代码执行,产出 `final/M-Gate-Report-v2.2.12.json`(标准 JSON 格式)。

Hidden Instructions

High
Category
Prompt Injection
Confidence
70% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · references/dispatch/G14-中文AI痕迹检测器.md (reported line 11)May include surrounding context.

md
> 公共工具白名单 / 零 exec / 叶子纪律 / 降级自报 / token 统计:见 [`_shared/真源/dispatch-header.md`](../_shared/真源/dispatch-header.md)

<!-- generated: dispatch-contract (do not edit by hand) -->
node: g14_style_gate
phase: Phase 4.4 前置(G14 风格闸)
default: required

Hidden Instructions

High
Category
Prompt Injection
Confidence
70% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · references/dispatch/G14-中文AI痕迹检测器.md (reported line 11)May include surrounding context.

md
> 公共工具白名单 / 零 exec / 叶子纪律 / 降级自报 / token 统计:见 [`_shared/真源/dispatch-header.md`](../_shared/真源/dispatch-header.md)

<!-- generated: dispatch-contract (do not edit by hand) -->
node: g14_style_gate
phase: Phase 4.4 前置(G14 风格闸)
default: required

Hidden Instructions

High
Category
Prompt Injection
Confidence
70% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · references/dispatch/T1b-定向回查.md (reported line 10)May include surrounding context.

md
> 公共工具白名单 / 零 exec / 叶子纪律 / 降级自报 / token 统计:见 [`_shared/真源/dispatch-header.md`](../_shared/真源/dispatch-header.md)

<!-- generated: dispatch-contract (do not edit by hand) -->
node: t1b_targeted_review
phase: Phase 4.3 定向回查(T1b)
default: required

Hidden Instructions

High
Category
Prompt Injection
Confidence
70% confidence
Finding

Hidden instructions were detected in comments or invisible text. These could contain malicious directives. Manual review is recommended.

Content

Scanner excerpt · references/dispatch/T1b-定向回查.md (reported line 10)May include surrounding context.

md
> 公共工具白名单 / 零 exec / 叶子纪律 / 降级自报 / token 统计:见 [`_shared/真源/dispatch-header.md`](../_shared/真源/dispatch-header.md)

<!-- generated: dispatch-contract (do not edit by hand) -->
node: t1b_targeted_review
phase: Phase 4.3 定向回查(T1b)
default: required

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

L26 明确给 T1/T2/T3/T1b 的 allow_research 增加 web_search、web_fetch、tavily 及 academic_extra 工具族;但 L61 随后写“子代理只有 read/write/edit(检索档的检索工具),不得触碰以上任何编排工具”。该表述把子代理能力收窄成仿佛仅有读写编辑,与前面实际允许的网络/学术检索工具不一致,容易误导执行者对允许工具面的理解。

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

L44 规定心跳文件由 worker 自己写入磁盘,但同段又说明 T6/T7/T9/G14 因为只有 read 权限需由主控代写;而 L45 进一步把是否开工绑定到对该心跳文件写盘的显式同意。这与前文对只读角色“报告由主控落盘”的权限边界形成直接矛盾,文档同时声称是 worker 行为又承认这些角色不能写。

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The protocol explicitly forbids using session history or querying the original session, yet the required §8 fields instruct the operator to obtain the original controller session_key from sessions_history if visible. This internal contradiction can cause implementers or agents to violate the earlier safety boundary, leading to unauthorized cross-session metadata access, policy bypass, or confused-deputy behavior during takeover handling.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The document defines 🟢/🟡/🔴 earlier as data trust levels (A/B/C), but the completion checklist later says those icons are only for timeliness ratings. In a multi-agent evidence pipeline, this inconsistency can cause agents or auditors to misclassify low-trust evidence as merely stale/fresh, weakening source-quality controls and downstream review decisions.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The YAML content and operational notes are entirely written in Chinese, including workflow instructions and required behavior, with no indication that users may choose another language or locale. Under the stated policy, forcing a specific language without opt-in is a natural-language policy violation unless the locale constraint is explicitly justified.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The file states that if the target language includes Chinese, G14 must run, while purely foreign-language content is marked not applicable. This is a language-specific policy constraint, and the text does not offer user choice or document a compliance or regional justification for the enforced Chinese-only handling.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The precondition requires '目标语言' to already be written into the task brief, making language selection mandatory as part of the workflow. Nowhere in this file does it offer the user a choice of language or indicate that the language constraint is optional or user-driven, which is a natural-language locale policy concern.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The file’s natural-language content is entirely in Chinese and provides no indication that users may choose another language or locale. Under the stated policy, forcing a specific language without user opt-in is a natural-language policy violation unless the locale restriction is explicitly justified.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · references/_shared/真源/中文数据源集成.md (reported line 72)May include surrounding context.

md
- **API**:完全免费、**无需 API Key**、REST 接口
- **中文文献覆盖**:收录大量中文学术期刊/学位论文(通过 Crossref/DOI 及中文出版商数据源)
- **能力**:按标题/作者/机构/主题/引用检索,支持全文搜索、过滤、分页
- **用法**:`GET https://api.openalex.org/works?search=人工智能&per-page=20`
- **适合**:文献计量、开题调研、引用分析

#### Crossref(强烈推荐)

External Transmission

Medium
Category
Data Exfiltration
Confidence
50% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · references/_shared/真源/中文数据源集成.md (reported line 79)May include surrounding context.

md
- **性质**:DOI 注册机构([api.crossref.org](https://api.crossref.org)),全球文献元数据
- **API**:免费、无 Key、REST 接口
- **能力**:按 DOI/标题/作者检索论文元数据
- **用法**:`GET https://api.crossref.org/works?query=深度学习&rows=20`
- **适合**:查 DOI、找英文文献元数据、引用格式

> **以上两者默认关闭**,须主人在 Phase 0 **显式勾选**才调用;未勾选或选「④全部拒绝」则不调用。

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The document contains contradictory operational guidance: earlier sections instruct the controller to read host-visible metadata such as session_status and perform limited Phase 0 availability checks, while lines 73–74 claim the skill runtime does not read host configuration or probe models. In a multi-agent orchestration skill, this inconsistency can cause reviewers, operators, or downstream components to misunderstand what environment data is accessed and when, undermining trust boundaries and potentially bypassing expected disclosure or approval paths.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

This is a markdown file, so vague-trigger review applies. The phrase "每个项目必跑" is extremely broad and does not specify project type, context, or exclusion conditions, which could cause unintended invocation across unrelated workflows.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

文件前文多处明确说明自 v2.15.13 起,T7 不再产出反哺报告,反哺报告改由主控在 Phase 5 后置撰写。但 L177 仍写“审计员交付审计报告 + 反哺报告时”,这会直接误导执行者认为 T7 需要同时提交两类产物,属于文档意图与实际流程定义相冲突。

Content

No source excerpt is available for this finding.

Ae4

Medium
Category
analysis-evasion
Confidence
80% confidence
Finding

Suspicious Unicode normalization or mixed-script content

Content

No source excerpt is available for this finding.

Ae4

Medium
Category
analysis-evasion
Confidence
80% confidence
Finding

Suspicious Unicode normalization or mixed-script content

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Low
Category
Not specified by scanner
Confidence
96% confidence
Finding

The file states that the historical name "M 机械化硬门" is deprecated and does not match the implementation, which is actually "主控 LLM 推理模拟执行" with zero shell/exec. This is a direct documentation-level contradiction between prior intent-labeling and actual behavior, fitting intent-code/documentation divergence.

Content

No source excerpt is available for this finding.