Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    AI Feature Safety Review

    1

    A pedantic, breadth-first safety review of the **AI surface** of a web application - chat, copilots, RAG pipelines, tool-using agents, MCP servers and model-generated content.

    $6

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    AI Feature Safety Review

    AI Feature Safety Review

    Example session with this skill installed

    Run an AI feature safety review on this repo. It's a Next.js app with a support chat built on the OpenAI SDK, RAG search over our documentation via pgvector, and an agent tool that can create Linear tickets from chat. Ingested docs come from user uploads. Tell me what an attacker could do and what to fix before launch.

    • Read your context and instructions
    • Compiled the ai feature safety
    • Generated the data export

    Summary

    Reviewed the AI surface of your Next.js app: 4 model call sites, 1 RAG pipeline, 3 tools, 2 agent loops. Sweep covered 86 checks - 61 PASS, 14 FAIL, 8 N/A, 3 UNKNOWN (see coverage.csv).

    Top findings

    SevIDFindingLocation
    CriticalTL-04Agent createTicket executes with a service-role API key; no per-user authorization - attacker-controlled doc content can steer ticket creation into any projectsrc/agent/tools.ts:41
    HighINJ-02RAG chunks from user uploads enter the prompt unlabeled; forged instructions in a PDF reach the chat answer and the Linear toolsrc/rag/ingest.ts:88
    HighOUT-01Model answers rendered as raw markdown without sanitization - stored XSS via an ingested doccomponents/ChatMessage.tsx:23
    MediumCOST-03No per-user token/spend cap on the chat endpointapi/chat/route.ts:57
    MediumRAG-05Vector store namespace not scoped by tenant; similarity search crosses tenants on shared termssrc/rag/store.ts:112

    Autonomy tiers found

    • Chat answer: T1 - Doc search: T2 - Linear ticket creation: T3 (reversible write, no approval, no idempotency)

    Next steps

    1. Execute createTicket with the requesting user's identity and a server-side authorization check per call (threat-model.md, blast-radius matrix)
    2. Label retrieved chunks as untrusted; sanitize markdown at the sink
    3. Add a spend cap and an eval set covering benign injection cases (eval-gaps.md)

    Full report: ai-safety/REPORT.md - coverage ledger: ai-safety/coverage.csv

    coverage.csv

    CSV · data export

    Generated

    Example file from a real run - the skill writes it into your workspace.

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

    About this skill

    A pedantic, breadth-first safety review of the AI surface of a web application - chat, copilots, RAG pipelines, tool-using agents, MCP servers and model-generated content. It finds the paths by which untrusted content becomes privileged action or data exposure, and the missing engineering around the model (limits, evals, observability, rollback). Not a jailbreak contest, not a red-team exercise.

    What you get

    • ai-safety/REPORT.md - findings with blast-radius severity, exact file:line locations, guardrail fixes and confidence
    • threat-model.md - AI surface inventory, data-flow traces and a tool ledger with autonomy tiers (T0-T4)
    • coverage.csv + findings.json - every check marked PASS / FAIL / N/A / UNKNOWN, machine-readable
    • eval-gaps.md - what your regression evals cover, what they cannot catch, what to add before the next prompt or model change
    • A 20-line chat summary: counts by severity, top 5 attack paths, autonomy tiers found

    How it works

    1. Inventory - every call site, prompt channel, tool, agent loop, retriever, memory store and output sink, labeled trusted/untrusted, with an autonomy tier per flow
    2. Blast-radius matrix - per reachable tool: whose identity executes it, where authorization is enforced, argument validation, reversibility, worst case if arguments are attacker-controlled
    3. Checklist sweep - mapped to OWASP LLM Top 10 (2026) and the OWASP Agentic Top 10: prompt injection via untrusted content, excessive agency, tool authorization and confused-deputy access, cross-tenant retrieval and memory leaks, sensitive data sent to providers, hidden-context exposure, unsafe output handling, unbounded cost, missing evals and kill switches
    4. Per-call-site cards - 12 questions per call site: prompt channels, attacker-writable inputs, validation at sinks, bounds, caching, logging, fail-open vs fail-closed, eval coverage
    5. Report - false-positive guards applied (a finding requires a path: untrusted channel -> capability -> impact), severity rubric, coverage ledger

    Guardrails

    Static-first: default mode needs no model calls and no API keys. Canary tests run only in a test environment with synthetic data and an explicit call/spend budget you approve. No jailbreaks against production, no data extraction, no mutations of vector stores, memory or real systems. Secrets and PII are redacted, never pasted wholesale. Your code is not changed - patches are proposed, applied only on request.

    Requirements

    • Repo access to a web app with AI features
    • ripgrep for the bundled static inventory script (recommended; the agent can also grep manually)
    • Nothing to install otherwise - works with any SKILL.md-compatible coding agent

    Known limitations

    • Dynamic checks need a test environment you approve; production black-box mode has severe coverage limits
    • Findings require a traced path - "prompt injection is possible" alone is not a finding
    • Non-AI authorization defects (plain IDOR etc.) are deduplicated and cross-referenced to a web hygiene audit
    • The review proposes fixes but does not change your code unless asked

    Static-first: dynamic canary tests need a test environment with synthetic data and your explicit approval - production red-teaming is out of scope by design. Black-box mode has severe coverage limits. A finding requires a traced path; prompt injection being possible alone is not reported as a vulnerability. Non-AI authorization defects are cross-referenced to a web hygiene audit rather than duplicated. The review proposes fixes but does not change your code unless asked.

    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 today

    Requires

    ripgrep

    Permissions

    Terminal / Shell
    Read Files
    Write Files

    File Scopes

    ai-safety/**

    Reads your repo and writes findings to ./ai-safety/. The bundled inventory script needs ripgrep and runs locally with no network access. Optional grey-box canary checks run only in a test environment you approve, within a call and spend budget you set. No model calls, no API keys required.

    Listedtoday

    What's inside

    Frequently Asked Questions

    No. The default white-box mode is purely static: it reads your repo and runs a local ripgrep inventory script. No model calls, no API keys, no network. Optional grey-box canary checks use a test environment you approve and stay within a call and spend budget you set.

    No. It is non-intrusive by design: no jailbreaks against production, no exploitation, no data extraction, and no mutations of vector stores, agent memory or real systems. Evidence comes from reading code and tracing data flows; dynamic checks are limited to benign canaries in a test environment.

    OWASP LLM Top 10 (2026 edition) and the OWASP Top 10 for Agentic Applications (ASI01-ASI10). The skill verifies the current editions at run time and states which versions the report uses, rather than hardcoding a taxonomy.

    It proposes exact guardrail fixes with file:line locations but does not modify your code unless you ask. The report distinguishes confirmed traced paths from probable ones and includes a coverage ledger showing what was and was not checked.

    New in Security