📦

rsi-skill

Name: rsi-skill
Author: lambdasec

Ревизия содержимого r1 Средний риск 📁 Доступ к файловой системе⚙️ Внешние команды🌐 Доступ к сети🔑 Переменные окружения

Evaluate and Improve Prompts With Test Cases

Prompt changes can look useful without improving real results. This skill scores examples, proposes revisions, and retains candidates that improve training scores without lowering validation scores.

Поддерживает: Claude Codex Code(CC)

📊 72 Адекватно

Maintain this Skill? Claim it 📊 72 Адекватно

Установить с помощью моего Агента

Скопируйте этот запрос в своего Агента. Он содержит каноническую страницу Skill и манифест.

Запрос агента

Review the Skillstore skill "rsi-skill" from https://skillstore.io/skills/lambdasec-rsi-skill.md and its manifest at https://skillstore.io/api/skills/lambdasec-rsi-skill/manifest. Verify the artifact. You may proceed after verification, subject to the environment's own policy.

Ваш Агент по-прежнему должен показать план и запросить все подтверждения, требуемые политикой безопасности.

Ресурсы для AI-агентов

Используйте эти ссылки, когда AI-агенту, crawler или script нужен чистый контекст вместо полной страницы.

Markdown-детали GET /skills/lambdasec-rsi-skill.md Подписанный manifest GET /api/skills/lambdasec-rsi-skill/manifest Подписанный lockfile GET /api/skills/lambdasec-rsi-skill/lockfile

Протестировать

Использование «rsi-skill». Summarize the documented GSM8K optimization example.

Ожидаемый результат:

The bundled example reports a baseline of 5/28 training cases and 2/12 validation cases.
The retained final revision reached 27/28 training cases and 11/12 validation cases.
Two intermediate revisions were discarded. These are documented results from one run, not guaranteed outcomes.

Использование «rsi-skill». Initialize a new sentiment classification task.

Ожидаемый результат:

Created a starter prompt for positive, negative, and neutral sentiment.
Added three labeled examples for editing.
This starter set is too small for validation splitting. Add representative cases before optimization.

Аудит безопасности

Средний риск

v1 • 28.10.2026 Открыть версионный отчет

Most static alerts concern harmless formatting, arithmetic examples, documentation links, or intended backend configuration. Four HTTP request findings expose credentials through unrestricted redirects; a separate model-grading weakness can undermine evaluation integrity. No evidence found of malicious instructions, embedded credentials, or intentional data theft.

Просканировано файлов

520

Проанализировано строк

Пункты проверки

Ложные срабатывания проигнорированы

Подтверждённые проблемы безопасности (1)

Средний

scripts/rsi.py:147-154 scripts/rsi.py:272-277

Untrusted content can influence model-based grading

Judge mode concatenates task input and generated answers into a grading message. Attacker-controlled content can request PASS verdicts and undermine prompt selection.

The grader receives untrusted answer text without explicit data-only instructions and accepts a PASS prefix. Exploit success depends on the selected model.

Пункты проверки возможностей (4)

Это реальные локальные возможности, которые могут ожидаться для этого навыка, поэтому они требуют проверки, но не считаются подтверждённым вредоносным поведением.

Высокий

scripts/rsi.py:41

Python HTTP libraries

req = urllib.request.Request(url, headers={"Authorization": f"Bearer {key}"} if key else {})

Authorization is attached as an ordinary request header before default urllib redirect handling. Cross-origin redirects can forward this bearer credential to another host.

Высокий

scripts/rsi.py:42

Python HTTP libraries

with urllib.request.urlopen(req, timeout=timeout) as r:

Default urlopen follows redirects without restricting the credential-bearing request to its original origin. A local redirect-handler check confirmed Authorization forwarding.

Высокий

scripts/rsi.py:96

Python HTTP libraries

req = urllib.request.Request(be["base"] + "/chat/completions", data=body, headers=headers)

The completion request includes Authorization in ordinary headers. A redirect accepted by urllib can send that credential to a different origin.

Высокий

scripts/rsi.py:97

Python HTTP libraries

with urllib.request.urlopen(req, timeout=300) as r:

The completion request uses default redirect handling without an origin check. Accepted redirects can carry the bearer credential to an unintended host.

Факторы риска

📁 Доступ к файловой системе (3)

references/example-gsm8k.md:5 references/example-gsm8k.md:43 scripts/rsi.py:80

⚙️ Внешние команды (28)

🌐 Доступ к сети (19)

scripts/rsi.py:30 scripts/rsi.py:41 scripts/rsi.py:42 scripts/rsi.py:96 scripts/rsi.py:97 scripts/rsi.py:33 scripts/rsi.py:34 scripts/rsi.py:76 scripts/rsi.py:80 scripts/rsi.py:12 scripts/rsi.py:34 scripts/rsi.py:81 SKILL.md:24 SKILL.md:81 SKILL.md:82 SKILL.md:83 SKILL.md:84 SKILL.md:85 SKILL.md:25

🔑 Переменные окружения (12)

scripts/rsi.py:65 scripts/rsi.py:68 scripts/rsi.py:75 scripts/rsi.py:76 scripts/rsi.py:77 scripts/rsi.py:12 scripts/rsi.py:65 scripts/rsi.py:75 scripts/rsi.py:77 scripts/rsi.py:78 scripts/rsi.py:82 SKILL.md:26

Аудитор:: codex

Поделиться и цитировать этот отчет

Делитесь версионным отчетом об оценке, нейтральным значком, встраиваемой карточкой и цитатами. Skillstore публикует доказательства, не решая, безопасен ли этот Skill.

Открыть версионный отчет

Копировать ссылку на отчёт

https://skillstore.io/skills/lambdasec-rsi-skill/audits/1?utm_source=security_passport&utm_medium=share&utm_campaign=versioned_report

Значок Markdown

[![Skillstore security assessment](https://skillstore.io/badges/skills/lambdasec-rsi-skill/security.svg)](https://skillstore.io/skills/lambdasec-rsi-skill?utm_source=security_passport_badge)

Значок HTML

<a href="https://skillstore.io/skills/lambdasec-rsi-skill?utm_source=security_passport_badge"><img src="https://skillstore.io/badges/skills/lambdasec-rsi-skill/security.svg" alt="Skillstore security assessment" loading="lazy"></a>

Встраиваемая карточка

<iframe src="https://skillstore.io/embed/skills/lambdasec-rsi-skill.html" title="Skillstore Security Assessment" sandbox="allow-popups allow-popups-to-escape-sandbox" loading="lazy" referrerpolicy="no-referrer" width="420" height="180"></iframe>

Академические ссылки (APA · BibTeX · CFF)

Цитата APA

lambdasec. (2026). rsi-skill security audit report (audit version 1) [Author version unspecified]. Skillstore. https://skillstore.io/skills/lambdasec-rsi-skill/audits/1

Цитата BibTeX

@techreport{lambdasec-lambdasec-rsi-skill-2026,
  author = {lambdasec},
  title = {rsi-skill security audit report (audit version 1)},
  institution = {Skillstore},
  year = {2026},
  number = {1},
  url = {https://skillstore.io/skills/lambdasec-rsi-skill/audits/1},
  note = {Author version unspecified}
}

CITATION.cff

cff-version: 1.2.0
message: "If you use this Skill, cite its author and this versioned security audit report."
title: "rsi-skill security audit report (audit version 1)"
version: "unspecified"
type: report
authors:
  - name: "lambdasec"
date-released: "2026-10-28"
url: "https://skillstore.io/skills/lambdasec-rsi-skill/audits/1"
identifiers:
  - type: other
    value: "skillstore:lambdasec-rsi-skill:audit:1"
    description: "Skillstore immutable audit report identifier"

Оценка Skillstore

Почему такая оценка Достоверность доказательств: Средний

Архитектура

Сопровождаемость

Контент

Сообщество

Соответствие спецификации

Что вы можете построить

Refine classification instructions

Compare a classifier prompt against labeled examples and inspect failures before accepting a revision.

Compare reasoning prompts

Use the bundled math cases to compare prompt variants with numeric checks and a reproducible validation split.

Test agent instruction changes

Evaluate instruction text against representative cases and review the proposed changes before applying them.

Попробуйте эти промпты

Create a first evaluation

Help me evaluate this prompt: [prompt]. Initialize a task directory and help me define expected answers. Confirm the backend before sending data.

Measure a baseline

Evaluate the prompt and cases in [directory] using [approved endpoint and model]. Report the baseline score, failed cases, and scoring methods without revising anything.

Run controlled optimization

Optimize [directory] for up to five rounds using [approved endpoint and model]. Keep the default validation split and copy rejection. Report scores and proposed changes.

Validate an instruction revision

Optimize [directory] using [evaluation model] and [proposal model] on [approved endpoint]. Reevaluate the best prompt on [independent test directory]. Review failures before requesting deployment approval.

Лучшие практики

Confirm the resolved endpoint and model before inference. Obtain approval for private remote data and avoid untrusted redirects.
Use representative cases and deterministic checks where possible. Reserve a separate final test set beyond the reused validation split.
Review failures and prompt changes, repeat evaluation, and obtain approval before replacing production instructions.

Избегать

Sending confidential prompts or cases to an unapproved remote endpoint.
Treating validation gains from a small, repeatedly reused split as proof of general improvement.
Using model-judge verdicts as the only acceptance gate for adversarial or security-sensitive tasks.

Часто задаваемые вопросы

Does this skill train model weights?

No. It revises instruction text and measures the resulting answers against test cases.

Which model services can it use?

It supports OpenAI-compatible chat endpoints, including OptiQ Cloud, local OptiQ servers, and explicitly configured services. API features vary by backend.

Can I keep evaluation data local?

Yes, with an approved local endpoint. Select it explicitly and verify backend resolution because cloud configuration takes priority over automatic local discovery.

How many cases should I provide?

The instructions recommend 10 to 40 representative cases. Fewer than six cases disable validation splitting.

How does it select a revision?

Training scores must improve, and validation scores must not decrease when a validation set exists. Complete copied test inputs are rejected by default.

Are improvements guaranteed and automatically deployed?

No. Results depend on the model, checks, and examples. The instructions require user approval before applying the saved best prompt.