Back to skill

Security audit

fund-cio-copilot

Security checks for vulnerabilities and agentic risk

Overview

This looks like a legitimate investment-copilot skill, but it should be reviewed before installation because it automatically stores sensitive fund records and can create external documents with limited per-action safeguards.

Review this before installing in a real fund workspace. Use it only in an isolated workspace with appropriate file permissions, treat the generated JSONL and Markdown memory files as confidential records, require approval before Tencent Docs publication or scheduled automation, and avoid promoting raw web or project text into memory without sanitization.

Vulnerability Patterns
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • 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
Findings (2)

T09 · Insecure Skill Coding Practices

Warning
Location
decision-memory.md:11
Finding

Confidential investment records are stored in plaintext workspace files

Content
View full analysis

Vulnerability Details

File Location: decision-memory.md, lines 11–14 and 124–126
Vulnerability Type: Plaintext storage of confidential decision and portfolio data
Risk Level: Medium

Vulnerable Snippet

The following is an English rendering of the complete relevant Markdown segment:

markdown
## Storage Scheme
- Format: **JSONL** (one Object per line: Decision / Monitoring / Outcome / Learning), directly queryable by code and readable by the LLM for Recall
- Path: `decision_memory.jsonl` (workspace root) + `outcome_memory.jsonl` (Outcome/Learning, associated by decision_id)
- Phase 1: zero external dependencies; linear scanning is sufficient

## Privacy Boundary
- Decision / Outcome / Learning Memory contains confidential fund data, is stored locally, is not externally shared, and is not used for third-party training
- Exporting or sharing requires manual approval from the IC secretary

Related persistence instructions also appear in SKILL.md, lines 187–190, and the sensitive record structures are defined in decision-object.md, lines 158–207.

Technical Analysis

The Skill directs the agent to persist investment decisions, evidence, valuation ranges, internal IC resolutions, portfolio financial results, and institutional learning records in JSONL files located at the workspace root.

Although the documentation states that these records are confidential and must not be externally shared, it does not define or enforce:

  • Owner-only filesystem permissions.
  • Encryption at rest.
  • Per-fund or per-user workspace isolation.
  • Retention and secure-deletion policies.
  • Exclusion from generic workspace synchronization, backup, indexing, or presentation tools.
  • Integrity protection against unauthorized modification.
  • Access controls for other agents or tools operating in the same workspace.

The Outcome Object and Learning Object may contain actual revenue, customers, gross margin, valuation, failed assumptions, misleading evide ...[truncated 1738 chars]

Remediation
View remediation

Remediation Suggestions

  1. Move decision memory into a dedicated, access-controlled storage directory rather than the workspace root.
  2. Create files with owner-only permissions, such as mode 0600, and directories with mode 0700.
  3. Encrypt sensitive records at rest using a managed key or operating-system credential store; do not store the encryption key beside the data.
  4. Separate records by fund, tenant, and authorized user to prevent cross-context access.
  5. Define an explicit retention schedule and secure-deletion procedure for expired projects and outcomes.
  6. Exclude the memory directory from generic file-delivery, synchronization, search-indexing, source-control, and backup mechanisms unless explicitly approved.
  7. Add integrity protection, such as authenticated encryption or signed records, to detect unauthorized changes.
  8. Minimize stored data by replacing unnecessary names and free-form confidential text with identifiers or redacted summaries.
  9. Enforce authorization checks before Recall, export, or sharing rather than relying only on a written approval rule.
  10. Record access and export events in a separate tamper-resistant audit log.

T02 · Agent Memory Poisoning

Warning
Location
mode-weekly-radar.md:31
Finding

Untrusted web and project content can be automatically persisted and recalled without instruction sanitization

Content
View full analysis

Vulnerability Details

File Location: mode-weekly-radar.md, lines 31–36
Vulnerability Type: Persistent memory poisoning through untrusted retrieved content
Risk Level: Medium

Vulnerable Snippet

The following is an English rendering of the complete relevant Markdown segment:

markdown
1. Define the window: last Monday through Sunday, based on public disclosure time, allowing one or two boundary-day cases
2. Search: multidimensional coverage using WebSearch
3. Verify: use company-registration data to verify key state-owned investment targets
4. Draft: prepare the document according to the reporting standard
5. Deliver: write `Local_Industry_Fund_Investment_Observation_YYYY-MM-DD.md` to disk and use `present_files`
6. Benchmark capture: execute Capture automatically—select reusable patterns or strategic anchor cases from the week, incrementally write them into the benchmark case library, and Recall historical peers for comparison

Related behavior is defined in:

  • mode-weekly-radar.md, lines 129–135: automatically writing benchmark cases and recalling them in later Screening and Evaluation workflows.
  • decision-memory.md, lines 11–14 and 33–60: JSONL records may be read by the LLM and automatically recalled.
  • decision-object.md, lines 35–103: many persisted fields accept unrestricted strings, including claims, evidence, sources, reasons, actions, assumptions, and conflicts.

Technical Analysis

The Skill ingests content from public web searches, company information, business plans, management disclosures, and user-supplied project material. Selected content is automatically written into persistent JSONL or Markdown memory and may later be read by the language model during Recall.

The Skill contains useful evidence-quality and source-verification rules, but it does not define controls that distinguish informational content from executable agent instructions. In particular, it does not require the implementation to:

  • Treat ...[truncated 3209 chars]
Remediation
View remediation

Remediation Suggestions

  1. Declare all web results, business plans, company disclosures, recalled records, and imported documents to be untrusted data.
  2. Wrap retrieved content in explicit data-only delimiters and add a binding instruction that embedded directives must never be followed.
  3. Detect and quarantine instruction-like content, including requests to ignore rules, invoke tools, reveal secrets, modify memory, or change output requirements.
  4. Use strict typed schemas with maximum lengths, allowed character sets, and enumerated values wherever possible.
  5. Store immutable source excerpts separately from agent-authored claims, conclusions, and actions.
  6. Never copy raw source text into action-oriented fields such as next_action, reason, updates, or tool parameters.
  7. Require human approval before promoting web-derived records into persistent benchmark memory or higher memory tiers.
  8. Sanitize records again at Recall time; do not assume previously stored content is safe.
  9. Limit Recall to the minimum fields necessary for the current task and exclude raw source bodies by default.
  10. Prevent recalled content from initiating tools, file operations, publication, memory writes, or policy changes.
  11. Add provenance metadata, content hashes, creator identity, and timestamps to every stored record.
  12. Validate JSONL integrity before use and reject records modified outside the authorized capture process.
  13. Separate market-intelligence context from decision evidence at the implementation level, not only through written instructions.
  14. Add adversarial tests using webpages and business plans containing prompt-injection strings to verify that the agent quotes them as data and never executes them.
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 (17)

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 218)May include surrounding context.

md
ecast → Actual 链路**(Decision 的 assumption/valuation/return 即校准基线,缺失标"缺校准基线");`SKILL.md` 数据验证增 **Data Foundation**(Comparable 缺数据源时标"估算,缺可比数据",不编造数字)

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The trigger list includes broad phrases such as '周报', '上会', '复盘', and '跟踪' that can plausibly appear in ordinary conversation outside this investment-specific skill. In an agent-routing system, overly generic triggers can cause unintended activation of a high-authority workflow that reads local configuration, writes files, captures memory, or initiates scheduled/report-generation behavior without the user clearly intending to invoke this skill.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

This section explicitly permits autonomous local writes, Tencent Docs document creation, and scheduled Radar/Portfolio execution, but the warning is buried in a policy section rather than presented as a prominent user-facing notice at the point of capability use. If the skill is triggered unexpectedly or the user misunderstands its autonomy, it could create documents, persist data, or schedule recurring actions that expose sensitive investment information or create unwanted operational side effects.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The _note text is written entirely in Chinese and instructs required agent behavior, which can impose a specific language/locale expectation without indicating any user choice or opt-in. Under the policy, forcing a specific language without user selection is a natural-language policy violation.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

This markdown skill specifies all operational instructions, behaviors, and warnings in Chinese only, with no indication that users may choose another language or that the skill is intentionally limited to a Chinese-speaking context. That creates a natural-language policy concern because it effectively forces a specific language without user opt-in.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

This markdown file consistently presents the skill schema and operational guidance in Chinese, including required field descriptions and decision rules. Because it does not offer a language/locale option or explain that the skill is intentionally limited to a Chinese-speaking context, it creates a natural-language policy concern under the language-choice requirement.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The file is written entirely in Chinese with embedded English terms and provides no indication that users may choose another language or locale. Under the policy rule for natural-language violations, a skill that effectively requires a specific language without opt-in should be flagged unless the constraint is explicitly justified.

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 instructions, prompts, and templates are entirely in Chinese, which effectively forces a specific language for users and operators. Under the policy, this is a language/locale constraint that should either be optional for the user or clearly justified as region-specific.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The documentation narrows the agent's boundary to decision support only and explicitly says it will not automatically manage portfolios or execute exit actions, but later introduces automatic re-evaluation behavior. In an investment workflow, that contradiction can cause the agent to take materially consequential autonomous steps without clear operator consent, creating governance, authorization, and process-integrity risk.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The file states that automatic review only runs after the user explicitly establishes scheduling, yet it separately says any key change triggers an automatic evaluation rerun. This inconsistent control model could let the skill bypass the stated human-in-the-loop safeguard and perform unsanctioned analysis actions in a sensitive investment decision pipeline.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The skill is written entirely in Chinese and presents required outputs and operating instructions only in Chinese, with no indication that users may choose another language. Under the policy, forcing a specific language without user 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
94% confidence
Finding

This markdown file contains user-facing operational instructions only in Chinese and does not indicate that users may choose another language. Under the policy rule, forcing a specific language without opt-in can be a natural-language policy violation unless the locale constraint is clearly documented and justified.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

L015-L017 explicitly state this Radar mode does not write to Decision Memory and keeps its benchmark library separate. However, L102 and L135 instruct automatic maintenance of '.workbuddy/decision_memory/benchmark_cases.jsonl', which is a Decision Memory directory path and undermines the stated separation. This is an active contradiction between the declared boundary and the documented storage behavior.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
87% confidence
Finding

The trigger set includes broad terms like '周报' that may appear in ordinary user requests, increasing the chance of unintended mode activation. In this skill, accidental routing is more concerning because Radar can automatically generate files, overwrite existing artifacts, and push outputs to external destinations, so misrouting can produce unintended side effects rather than just a bad answer.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The mode explicitly instructs overwriting same-name files, writing to workspace root, updating benchmark stores, and creating new Tencent Docs documents, but it does not require a clear user-facing confirmation before modifying or publishing data. In an agent context, silent writes and external document creation are meaningful side effects that can surprise users, leak work product to third-party services, or destroy prior artifacts through overwrite behavior.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The entire skill is written to operate in Chinese, including headings, instructions, and required output conventions, with no indication that users may choose another language. This creates a natural-language policy concern when a skill implicitly mandates one language without opt-in or documented locale justification.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Low
Category
Not specified by scanner
Confidence
95% confidence
Finding

L079 says the report must not include a standalone 'AI生成提示' line, while L100 defines the final structure as including '文末信息说明 + AI生成提示'. These instructions point the implementation in opposite directions for the same output artifact, creating an internal intent/documentation conflict.

Content

No source excerpt is available for this finding.