agentleFS
Sign inStart free
thejefflarsonthejefflarson/soundcheck/.claude/skills/prompt-injection/SKILL.md

prompt-injection

Skill20 starsChanged 4 months agoMITView on GitHub

Detects LLM prompts built from user input or retrieved documents that let

8 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use threat-modeling.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • msewellSkill

    threat-modeling

    Produces structured threat models for software systems using STRIDE on data flow diagrams. Generates DFDs with trust boundaries, identifies threats per element, scores risks, and defines concrete mitigations. Outputs a complete threat model Markdown document through phased, interactive delivery. Use when threat modeling a system, analyzing security threats, identifying attack surfaces, performing STRIDE analysis, assessing security risks in architecture, creating a threat model document, or when the user mentions threat model, attack surface, trust boundaries, STRIDE, or security risk analysis.

    Also in 4 projects that use this

The file

We split this skill into the rules it gives an agent

2 rules. 0 are common, 2 are distinctive, and 0 are very specific to this project.

  • 0

    Common

    Most of the 7 edited versions we read keep these.

  • 2

    Distinctive

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    No structural separation between developer instructions and untrusted data — everything is one string.

    line 24
  2. Distinctive:

    Output is validated before any downstream action: Every code path that uses the model's response — returning it to the caller, rendering …

    line 42

“Common” is measured: kept by at least 6 in 10 of the 7 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Prompt Injection Security Check (OWASP LLM01:2025)

What this checks

Protects against attacker-controlled text that hijacks LLM instructions. Direct injection arrives through user input; indirect injection arrives through retrieved documents, emails, or tool outputs. Both can cause the model to exfiltrate data, bypass guardrails, or execute unintended actions.

Vulnerable patterns

  • User input interpolated directly into the system-role message — user text lands in the instruction tier.
  • Retrieved documents concatenated raw into the prompt with no delimiter or trust label.
  • Email bodies, fetched web pages, or other external content passed into the prompt with no boundary markers separating data from instructions.
  • No structural separation between developer instructions and untrusted data — everything is one string.
  • Raw model response returned to the caller, rendered, logged, or used to trigger a downstream action with no validation step in between.

Fix immediately

Flag the vulnerable code and explain the risk. Then suggest a fix that establishes these properties. Translate each property into the audited file's language and LLM client library — use that library's documented role-separated message API rather than mirroring an example from another stack.

  1. Trust tiers are structurally separate. Developer instructions go in the system role; user input and retrieved documents go in the user role, wrapped in explicit delimiter tags that label the content as untrusted data. Never interpolate user text into the system prompt.
  2. Input is bounded and screened before the API call. Apply a length cap and reject obvious injection markers (phrases like "ignore previous", "new instruction"). Screening is a denylist and will not catch everything, but it raises the bar.
  3. Output is validated before any downstream action. Every code path that uses the model's response — returning it to the caller, rendering it, logging it, triggering a tool call — first routes it through a gate that enforces size bounds and rejects suspicious instruction language. A defined validator that is never called does not satisfy this.

Verification

Confirm these properties hold:

  • User input never appears in the system-role message
  • Retrieved documents are wrapped in explicit delimiter tags, not concatenated raw into the prompt
  • Input length and pattern validation runs before the API call
  • The LLM response passes through a validation step at every call site before it is returned, rendered, logged, or used to trigger an action. A validator defined but never invoked does not count

References

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

Check your setup, free