More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    LLM App Security Triage

    1

    An application that passes user text to a model has an attack surface that no traditional scanner covers.

    $10

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    LLM App Security Triage

    LLM App Security Triage

    Example session with this skill installed

    We added a support assistant that reads user tickets and can look up orders through a tool call. Before we ship, triage the llm-assistant/ folder for the LLM-specific problems and tell me where to look first.

    • Read your context and instructions
    • Compiled the llm app security

    LLM App Security Triage Report

    Scanner: llm-app-triage v1.0.0
    Target: llm-assistant (working tree)
    Files scanned: 10 | Findings: 21 | Files with findings: 6
    Highest severity: critical | Rules fired: 20 of 20 | Families: 6 of 6

    The scan read every text file it could parse and applied all 20 rules in the catalogue. 21 findings landed in 6 files. The other 4 files were silent, which is the point: the scanner is not tuned to shout at ordinary code.

    Findings

    RuleFamilySeverityFileLineWhat it matched
    MOD-001model output executioncriticalsrc/planner.py7Model output evaluated as code
    PI-001prompt injectioncriticaldocs/support-ticket.txt3Text that tries to discard earlier instructions
    PI-002prompt injectioncriticaldocs/support-ticket.txt5Text that tries to hide activity from the operator
    SEC-001secretscriticalconfig/settings.py3Cloud access key identifier
    SEC-002secretscriticalconfig/settings.py5Source-hosting access token
    SEC-003secretscriticalconfig/settings.py7Messaging platform token
    SHL-001shell constructioncriticalsrc/agent_tools.py24Shell interpreter flag enabled on a call
    SHL-002shell constructioncriticalsrc/agent_tools.py16Command handed to an interpreter call built from parts
    MOD-002model output executionhighsrc/planner.py11Model output handed to a process call
    PI-003prompt injectionhighdocs/support-ticket.txt7Text that tries to extract the model instructions
    PI-004prompt injectionhighdocs/support-ticket.txt9Text that tries to put the model into an unrestricted mode
    SEC-004secretshighconfig/settings.py9Credential assigned to a literal value
    SHL-003shell constructionhighsrc/agent_tools.py12Command built with an interpolated value
    SHL-004shell constructionhighsrc/agent_tools.py20Interpreter call given a bare variable
    SQL-001query constructionhighsrc/reporting.py7Query keyword concatenated with a value
    SQL-002query constructionhighsrc/reporting.py11Query keyword built with an interpolated value
    SQL-003query constructionhighsrc/reporting.py15Query keyword built with percent formatting
    TAI-001untrusted data in the prompthighsrc/retrieval.py12Untrusted-looking value concatenated into a prompt
    TAI-002untrusted data in the promptmediumsrc/retrieval.py17Prompt built with an interpolated value
    TAI-003untrusted data in the promptmediumsrc/retrieval.py7Template placeholder filled from user input
    TAI-003untrusted data in the promptmediumsrc/retrieval.py22Template placeholder filled from user input

    Families that fired

    FamilyRulesFindingsHighest severity
    prompt injectionPI-001, PI-002, PI-003, PI-0044critical
    secretsSEC-001, SEC-002, SEC-003, SEC-0044critical
    shell constructionSHL-001, SHL-002, SHL-003, SHL-0044critical
    query constructionSQL-001, SQL-002, SQL-0033high
    untrusted data in the promptTAI-001, TAI-002, TAI-0034high
    model output executionMOD-001, MOD-0022critical

    Severity breakdown

    SeverityFindings
    critical8
    high10
    medium3

    How the evidence is shown

    A credential finding never echoes the value. The scanner reports credential-shaped value on this line and points at the file and the line, so the report can be pasted into a ticket without leaking the key a second time. An injection finding reports instruction phrasing detected for the same reason: quoting the hostile phrase back would copy it into your ticket, your log and your chat history.

    The shell, query, taint and model families quote the source line, because in those families the construct itself is the finding and it carries no key material.

    Recommendation

    Start with the two model-output findings and the four injection findings, because those six are the ones where text changes behaviour instead of merely looking wrong. MOD-001 is the worst line in the tree: a model response handed straight to an interpreter turns a hallucination into code execution. PI-001 and PI-002 sit in a support ticket, which means the hostile text arrived through the retrieval path and would have been inside the context window. The four secret findings in config/settings.py are rotation items rather than cleanup items, because a key that reached the remote is already compromised.

    The shell and query findings are the ordinary part of an LLM backend and they are still real: a command assembled from parts and a query assembled by concatenation are code execution and injection waiting for a reachable value. The three taint findings are the structural defect underneath all of it, because user input is concatenated straight into a prompt with nothing separating it from the instructions.

    Every finding carries a rule ID, a file and a line, so each one is a ticket rather than a paragraph, and the rule identifiers are stable, so the next scan can be compared against this one.

    Next steps

    1. Remove the eval on the model response and parse it against a fixed schema instead.
    2. Treat every document that reaches the model as untrusted input: keep it in a delimited block and say that it is data.
    3. Rotate the four credentials at their issuers, then load them from the environment or a secret manager at run time.
    4. Replace the concatenated and formatted queries with parameterised statements.
    5. Rebuild the command calls as argument lists and turn the shell interpreter flag off.
    6. Separate user input from the instructions in the prompt builder, then re-run the triage and confirm the findings are gone.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    An application that passes user text to a model has an attack surface that no traditional scanner covers. The request can try to redirect the model, the retrieval index can contain poisoned documents, a tool call can be steered to an unsafe operation, and the model's answer can be rendered back to a browser without escaping.

    This skill triages an LLM-backed application against twenty rules in six families: prompt injection, secrets, shell construction, query construction, untrusted data reaching the prompt, and model output reaching an interpreter. It works over source files, over logs, and over a single sample of model input, which makes it usable in three different places: before a release, during an incident, and when you want to understand one captured prompt.

    The analyser is pattern based and deterministic. It does not call a model, it does not need an API key, and it does not send anything anywhere. It is a first pass that tells you where to look, with stable rule identifiers so two runs can be compared.

    The analyser reads one file at a time, so it cannot follow a value from one file to another - that trace is a manual step. Pattern matching catches known shapes, not novel attacks. It does not call a model or your provider API, and it does not execute the code under review.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 1 day ago

    Requires

    Python 3.9 or newer

    Permissions

    Terminal / Shell
    Read Files

    File Scopes

    **/*

    Shell is used to run the bundled analyser and read access is used to load the files you point it at. It writes no files, opens no network connections and reads no environment variables.

    Listed1 day ago

    What's inside

    Frequently Asked Questions

    No. The analysis is pattern based and runs locally. Nothing leaves the machine.

    A folder of source files, a single file, a captured log, or one sample of model input passed directly on the command line.

    No. It catches the known instruction shapes and the unsafe code paths around them. It is a triage tool, and a determined injection will use wording it does not know.

    No, and it should not be presented as one. It tells you which files deserve attention first.

    It works on text, so it applies to any language, but the rules are written for the shapes that show up in Python, JavaScript, TypeScript, Go and Java code bases.

    No. The analyser reads one file at a time, so it cannot tell whether an untrusted value in one file reaches the prompt in another. That trace is the part you do by hand, and the package includes the review procedure for it.

    New in Security