T08 · Insecure Dependencies
- Location
README.md:162- Finding
Unpinned installer and mutable remote Skill source create supply-chain exposure
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This is a disclosed markdown skill for quantitative-research workflows, but users should pin installs and avoid deploying generated trading or data-system code without review.
Install only from a pinned, reviewed commit or release. Treat generated code as design guidance, not deployable production infrastructure: add authentication, authorization, rate limiting, compliance review, and testing before using any trading, market-data, or execution-system snippets. Expect local state retention in state/research-context.md and remove it if it contains sensitive strategy information.
README.md:162Unpinned installer and mutable remote Skill source create supply-chain exposure
roles/09-bbg-data-pipeline.md:1553Generated market-data API omits enforced authentication and authorization
Referenced artifact was not completely inspected
| 06 | 因子模型 | AQR | roles/06-aqr-factor-builder.md | 02 | normal | 2 |
Referenced artifact was not completely inspected
| 12 | 组合优化 | Man Group | roles/12-man-portfolio-optimizer.md | 01,02,03,06 | normal | 2 |
The anomaly detector labels records as 'future timestamps' by comparing against a hardcoded 2024-01-02 reference instead of the real current time or event-time watermark. This can misclassify valid market data as invalid or miss truly bad timestamps, undermining data quality controls and enabling silent pipeline corruption or denial of service through excessive false positives.
This PIT example claims to support as-of retrieval but the versioning/write pattern can return a later restated snapshot rather than the data actually visible at the decision time. In a quant research pipeline, that creates look-ahead bias and silently contaminates backtests, model validation, and risk estimates with future information.
The fencing-token write validation claims to accept writes only from the current valid leader, but it compares epochs without checking whether the presented token itself is expired. In an active-active or failover scenario, a stale former leader could continue writing during or after lease expiry, undermining split-brain protection and causing duplicate or conflicting trading state.
The README instructs users to run npx skills add ... without pinning the package version or otherwise constraining what code is fetched and executed. This creates a supply-chain risk: if the skills package or its dependencies are compromised or a breaking update is published, installation could execute unexpected code in the user's environment.
The instructions direct users to clone the repository into a persistent local skill directory and state that it should then activate automatically. Persisting an externally sourced skill across sessions increases risk because the code/prompts remain available for future invocation, including unintended or attacker-induced activation, and may continue to operate after the initial trust decision is forgotten.
> Install the `quant-research-lab` skill. Follow these steps:
>
> 1. Ensure `.claude/skills/` directory exists in the current project (create if not)
> 2. Run `git clone https://github.com/yili1992/quant-research-lab.git .claude/skills/quant-research-lab`
> 3. Verify: `ls .claude/skills/quant-research-lab/` should show `SKILL.md`, `roles/`, `pipelines/`
> 4. Confirm installation is complete. When I describe quantitative research needs, this skill should activate automatically.
The skill declares that it activates automatically for broad 'quantitative research needs', which lacks clear boundaries and can cause unintended invocation in unrelated conversations. Over-broad auto-activation increases the chance that the skill's prompts, behaviors, or file reads/writes are triggered without explicit user intent, expanding attack surface and enabling prompt-trigger abuse.
The natural-language trigger list includes generic phrases such as asking for a backtest, risk analysis, or using a named role, which are common in normal conversation and may invoke the skill accidentally. In an agent environment, generic triggers are dangerous because an attacker can embed them in surrounding content to coerce tool/skill activation and influence subsequent workflow execution.
The README instructs users to run npx skills add without pinning a specific package version or commit, which allows execution of whatever package/version is current at install time. In an agent-skill context this creates a supply-chain risk: a compromised upstream package, namespace takeover, or breaking update could lead to unreviewed code execution during installation.
The listed trigger examples include very generic phrases such as '做一下回测' that can appear in ordinary discussion, making activation boundaries unclear. In an autonomous agent environment, overly broad triggers can cause the skill to activate unexpectedly, load prompts, read state, or initiate workflows when the user did not intend to invoke this skill.
The '自然语言触发' section states that describing a quant research need will automatically activate the skill, but does not define strict boundaries or confirmation requirements. Because the skill can orchestrate multiple roles, persist state, and recommend downstream actions, ambiguous activation materially increases the risk of unintended execution and context consumption.
La commande d’installation npx skills add ... exécute un outil distant/non figé sans version épinglée, ce qui expose l’utilisateur à des changements non audités de la CLI ou à une compromission de la chaîne d’approvisionnement. Dans le contexte d’une documentation de skill, cela peut mener à l’exécution de code localement lors de l’installation.
L’énoncé selon lequel la compétence s’active automatiquement pour des “besoins de recherche quantitative” est trop large et ne définit pas de frontière opérationnelle claire. Dans ce contexte multi-agent, cette ambiguïté augmente le risque d’activation accidentelle avec chargement de prompts complexes et exécution de flux non sollicités.
La documentation décrit une activation par langage naturel avec des formulations larges qui peuvent correspondre à des demandes ordinaires et déclencher la skill involontairement. Une activation non intentionnelle est dangereuse ici car la skill orchestre plusieurs sous-agents, lit/écrit l’état local et peut pousser l’utilisateur vers des actions d’installation ou d’exécution non désirées.
Les exemples de mode boîte à outils incluent des expressions courantes et vagues qui peuvent apparaître dans des conversations normales, entraînant une invocation involontaire d’un rôle spécialisé. Comme la skill peut ensuite dispatcher un sous-agent et persister des sorties dans des fichiers d’état, l’impact dépasse une simple mauvaise UX et devient un risque de comportement non autorisé.
The README frames the skill as a quant research workflow for design, backtesting, risk management, and validation, and explicitly says it is for research and educational purposes only. However, later sections document execution-oriented roles and pipelines such as Execution Algorithm, Trading System, Market Making, and system deployment/compliance flows, which indicate operational trading support rather than only research use.
The README recommends installing and invoking tooling via npx skills from a remote source without pinning an exact package version or immutable reference. This creates a supply-chain risk: a later compromised or malicious package release could be fetched and executed implicitly during installation or use.
The skill is documented to auto-activate on broad quant-related requests, which can cause unintended invocation in response to ordinary discussion rather than explicit user consent. In a skill that can orchestrate multiple roles, write state, and guide execution workflows, overly broad triggers increase the chance of surprising behavior and privilege overreach.
Example phrases like general backtesting or quant requests are too unspecific to safely trigger a complex orchestration skill. Ambiguous triggers increase accidental activation risk, which is more concerning here because the skill persists outputs and can progress into execution and system-design flows.
Relying on ambiguous natural-language matching rather than a clearly bounded trigger set makes the skill easier to invoke unintentionally or via prompt confusion. Because this skill can dispatch sub-agents and persist detailed context, accidental activation has meaningful privacy and operational consequences.
The README states that results are saved to state/research-context.md but does not provide a clear, prominent warning about data retention. Users may unknowingly cause sensitive trading ideas, project context, or proprietary inputs to be stored persistently on disk, creating confidentiality and compliance risk.
The README explains that complete role outputs are appended to a persistent audit file, but this retention behavior is not surfaced as a prominent warning near installation and usage. In a quant-research context, these outputs may contain proprietary strategy logic, risk assumptions, market views, or other sensitive material, so silent persistence materially increases exposure.
Suspicious Unicode normalization or mixed-script content
The README instructs users to run npx skills add against a remote GitHub skill without any pinned commit, tag, checksum, or version constraint. This creates a supply-chain risk: the referenced package or fetched skill contents can change over time, causing users to install unexpected or malicious code or prompts.