security-agent-guard
/security-agent-guard
Przegląda kod z LLM i agentami pod kątem iniekcji promptów, nadmiernych uprawnień i wycieków w RAG, wskazując plik i linię.
Ocena i bezpieczeństwo
- Pliki w paczce
- 7
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 13.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Przegląda kod z LLM i agentami pod kątem iniekcji promptów, nadmiernych uprawnień i wycieków w RAG, wskazując plik i linię.
Oryginalny opis (EN)
Reviews code that uses LLMs or AI agents for AI-specific security risks - prompt injection, excessive agent permissions, unsafe tool calls, cross-user data leaks in retrieval (RAG), secrets or personal data in prompts, and unsafe use of model output. Reports evidence-based findings with file and line, severity, certainty, and a fix. Use when the user asks to review the security of an AI feature, chatbot, agent, tool integration, or RAG pipeline. Not for general code security or committed secrets.
Jak działa
- Purpose
- When to Use
- Prerequisites
- Instructions
- Report format
- Examples
Podgląd SKILL.md
--- name: security-agent-guard description: Reviews code that uses LLMs or AI agents for AI-specific security risks - prompt injection, excessive agent permissions, unsafe tool calls, cross-user data leaks in retrieval (RAG), secrets or personal data in prompts, and unsafe use of model output. Reports evidence-based findings with file and line, severity, certainty, and a fix. Use when the user asks to review the security of an AI feature, chatbot, agent, tool integration, or RAG pipeline. Not for general code security or committed secrets. license: MIT compatibility: Read-only. Needs access to the source code. Works with any language or framework; examples use Python and JavaScript. metadata: collection: vibe.md group: security author: NoaGaler --- # AgentGuard: AI application security review ## Purpose Find security problems that exist because an application uses a language model: the model reads untrusted text, chooses actions, calls tools, retrieves data, and produces output that other code trusts. Report only what the code shows, explain why it matters, and suggest a fix the developer can apply. ## When to Use - The user asks whether an AI feature, chatbot, agent, or RAG pipeline is secure. - The user asks for a security review of code that calls a model API, defines tools for a model, or builds prompts. - The user asks about prompt injection, agent permissions, or data leaks through an AI feature. Do not use this skill: - to review code that does not use a model. Use `pre-commit-security-review` if it is installed. - to find secrets committed to the repository. Use `secret-guard` if it is installed. - to build or change an AI feature. This skill only reviews. ## Prerequisites - Read access to the source code. - No tools, accounts, or network access are required. - Permissions: read-only. The skill reads files and, when available, runs `git status` and `git diff` to find changes. It never edits, deletes, commits, or runs project code. ## Instructions 1. **Set the scope.** Review what the user named. If they named nothing, review the current changes (`git diff --staged`, then `git diff`, plus new untracked files). If they ask for the whole project, start with files that use a model and say which files you did not reach. 2. **Find the model calls.** Search the scope for model SDKs and APIs, prompt templates, tool or function definitions, agent frameworks, MCP servers, and vector stores. If there are none, say so and stop: this skill does not apply. 3. **Map each AI feature.** For each model call or agent loop, write down: - **sources:** where text enters the prompt, and which of it is untrusted; - **capabilities:** tools the model can call and data it can see; - **sinks:** where the output and tool arguments end up. 4. **Load only the references that apply.** - Untrusted text reaches a prompt: [prompt injection](references/prompt-injection.md). - The model has tools or acts on its own: [excessive agency](references/excessive-agency.md) and [tool calling](references/tool-calling.md). - The code retrieves documents or embeddings into prompts: [RAG security](references/rag-security.md). - Prompts contain configuration or user data, or conversations are stored: [sensitive information](references/sensitive-information.md). - Output is used for anything other than plain text shown to the same user: [output handling](references/output-handling.md). 5. **Trace each path from source to sink.** Apply the "When it is a finding" and "When it is not a finding" rules in the reference. A finding needs a location in the code for every part of its path. 6. **Separate certainty from suspicion.** Mark each finding **confirmed** when the whole path is in code you read, or **suspected** when part of it is in code you could not read, and say which part and how to check it. Never present a suspicion as fact. 7. **Do not report** generic advice, missing best practices without a path to harm, or risks in code outside the scope. Do not print secret values. 8. **Write the report** in the format below, highest severity first. 9. **Do not change code** unless the user asks. If they ask, fix one finding at a time and show the change. ## Report format For each finding: - **Where:** file and line for each part of the path. - **What:** the problem in one sentence, with the risk name. - **Why it matters:** a concrete misuse, such as the input that would cause it. - **Severity:** high, medium, or low, with a one-line reason. - **Certainty:** confirmed or suspected. - **Fix:** the change to make, and why it is stronger than the alternatives. End with: - **Coverage:** which AI features and which risk areas were checked, and which were skipped as not applicable. A clean result covers only what was checked. - **Out of scope:** issues noticed that belong to other reviews, such as general web security or committed secrets, in one line each. ## Examples **User:** "Is my support chatbot secure? The code is in `support/`." **Agent:** Finds the model call and its tools in `support/`, maps sources (customer messages, order records), capabilities (`lookup_order`, `issue_refund`), and sinks (the reply shown in the staff dashboard). Loads the prompt injection, excessive agency, tool calling, and output handling references. Reports that `issue_refund` runs without confirmation for any amount the model provides (high, confirmed), and that replies are inserted with `innerHTML` in the dashboard (high, confirmed). Lists RAG as not applicable. **User:** "Review my changes before I commit." **Agent:** Reads the diff. The changes touch only CSS and a date helper, with no model calls. Says that this skill does not apply and suggests a general security review instead.
Pliki w paczce (7)
- SKILL.md 6 KB
- references/excessive-agency.md 4 KB
- references/output-handling.md 4 KB
- references/prompt-injection.md 4 KB
- references/rag-security.md 4 KB
- references/sensitive-information.md 4 KB
- references/tool-calling.md 4 KB
Wersje
W bazie nie ma innych skilli o nazwie security-agent-guard.
Migawka bazy z 13.11.2026 · wersja paczki 551f7c4ff06e (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 90agent-security-audit
Audytuje konfiguracje agentów AI pod kątem nadmiernych uprawnień, powierzchni prompt injection, wycieków danych i brakujących zabezpieczeń. - 96security-review
Przegląda kod i zmiany pod kątem podatności: wstrzyknięć, sekretów, niebezpiecznego wejścia i uprawnień. - 90skill-security-auditor
Sprawdza skille, prompty i skrypty pod kątem wstrzyknięć promptu, wycieków sekretów, niebezpiecznych poleceń i nadmiernych uprawnień. - 100agent-security
Sprawdza kod LLM, RAG, MCP i agentów pod kątem problemów bezpieczeństwa z OWASP i recenzuje zmiany w systemach AI. - 100security-reviewer
Przegląda kod i projekt pod kątem wstrzyknięć, słabego uwierzytelniania, sekretów i iniekcji promptów, uzupełniając SAST i DAST. - 90common-llm-security
Sprawdza aplikacje AI pod kątem listy OWASP LLM Top 10 z 2025 roku, w tym promptów, narzędzi agentów i RAG.