From soundcheck
Flags prompt injection, SSRF, and context flooding in RAG pipelines ingesting docs or fetching URLs for LLM prompts. Recommends domain allowlists, truncation, delimiters, logging.
npx claudepluginhub thejefflarson/soundcheck --plugin soundcheckThis skill uses the workspace's default tool permissions.
Prevents prompt injection through retrieved documents and uncontrolled content flooding
Detects prompt injection vulnerabilities in LLM code constructing prompts from user input, system prompts, RAG pipelines, or external data. Suggests fixes with trust tiers, delimiters, input screening, and output validation.
Provides defense techniques against prompt injection attacks including direct, indirect injections, and jailbreaks, grounded in reference patterns, sharp edges, and validations. Use when LLM security terms mentioned.
Applies LangChain security best practices: secrets management, prompt injection defense, safe tool execution, and LLM output validation for production apps.
Share bugs, ideas, or general feedback.
Prevents prompt injection through retrieved documents and uncontrolled content flooding into LLM context. Attacker-controlled documents can override system instructions, exfiltrate data, or manipulate model behavior when injected without guardrails.
prompt = system_prompt + retrieved_doc — retrieved content can override instructionsrequests.get(user_url).text — arbitrary URL fetch with no domain allowlist (SSRF)Flag the vulnerable code and explain the risk. Then suggest a fix that establishes these properties:
ssrf skill for outbound HTTP.user role — never concatenated into
the system prompt. The model is more likely to treat it as data rather
than instructions when the framing is structural. See the prompt-injection
skill for the trust-tier pattern.Anchor — shape, not implementation:
require(host_of(url) in ALLOWED_DOMAINS)
doc = fetch(url, timeout=5)[:MAX_CHARS]
log_retrieval(url, len(doc))
prompt = {
system: DEV_INSTRUCTIONS,
user: f"<context untrusted>{doc}</context>\n<question>{query}</question>",
}
Confirm the response: