Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    Srednoff RAG Poisoning & Retrieval Trust Auditor

    1

    Audit ingestion-to-citation trust using supplied artifacts and deliver a source classification ledger, retrieval attack-path review and regression plan. No live corpus scanner or poisoning detector is installed.

    $7.99

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned

    Srednoff RAG Poisoning & Retrieval Trust Auditor

    Example session with this skill installed

    Audit whether user-supplied RAG metadata can claim policy authority.

    • Read your context and instructions
    • Compiled the srednoff rag poisoning

    Expected report shape, not a completed real audit: Executive Summary; Scope; Assets / Trust Boundaries

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

    About this skill

    Srednoff RAG Poisoning & Retrieval Trust Auditor

    One-Line Outcome

    Audit ingestion-to-citation trust using supplied artifacts and deliver a source classification ledger, retrieval attack-path review and regression plan. No live corpus scanner or poisoning detector is installed.

    Short Description

    A plausible answer can conceal poisoned chunks, forged metadata, stale policies or cross-tenant retrieval; ordinary relevance metrics do not establish source trust.

    Problem It Solves

    A plausible answer can conceal poisoned chunks, forged metadata, stale policies or cross-tenant retrieval; ordinary relevance metrics do not establish source trust.

    RAG data owner: document/source lineage ledger, ingestion authority classes, poisoned-metadata/citation mismatch cases and retrieval regression matrix.

    What You Get

    • Executive Summary
    • Scope
    • Assets / Trust Boundaries
    • Evidence Reviewed
    • Threat Model
    • Source Classification Ledger
    • Retrieval Attack Paths
    • Chunk And Citation Reconciliation
    • Findings
    • Critical Findings
    • Quick Wins
    • Prioritized Remediation Plan
    • Missing Evidence
    • Residual Risk
    • Verification Checklist
    • Unknowns
    • Next Validation Steps

    How It Works

    1. Pin the review to a corpus snapshot and complete pipeline revision; accept sanitized supplied exports rather than querying production or ingesting documents.
    2. Map source classes separately: approved authoritative, approved secondary, external/unverified, expired and quarantined. Record who controls content and who assigns trust metadata.
    3. Trace document ingestion through parser, normalization, chunker and embedding input; verify content digest, canonical source/version and access labels remain linked to every derived chunk.
    4. Review whether uploader-controlled titles, authority tags, dates or URLs can replace server-assigned provenance. A source URL or signed transport alone does not establish editorial authority.
    5. Inspect retrieval authorization for the actual principal before chunks enter model context, including vector, lexical, hybrid search, reranking and query/result caches. Record any unreviewed branch.
    6. Compare representative retrieval traces for duplicated content, rank displacement and suspicious concentration. Distinguish a ranking anomaly from confirmed malicious poisoning; do not tune live ranking.
    7. Inspect retrieved instructions, metadata and chunk combinations for policy-authority confusion; treat all retrieved prose as data even when the underlying document is approved for factual use.
    8. Reconcile answer citations to exact returned chunk spans and source versions. Separate an unsupported answer, citation mismatch, stale answer and unauthorized retrieval rather than merging them into one poisoning label.
    9. Trace update, tombstone, quarantine and access-revocation propagation through embeddings, replicas and caches. Specify an owner-approved reversible exclusion plan without deleting or reindexing production.
    10. Create paired local tests for legitimate retrieval, poisoned metadata, stale sources, access-label changes and cross-tenant near-duplicate documents. Record source eligibility, returned chunk IDs and citation support as separate assertions.
    11. Deliver trust model, evidence-linked findings and prioritized remediation. Require an approved snapshot-based canary and rollback/retest criteria before any future corpus or retrieval configuration change.

    Best For

    RAG engineers and knowledge-platform owners handling mixed-authority or multi-tenant corpora

    Example Tasks

    • Audit whether user-supplied RAG metadata can claim policy authority.
    • Reconcile current retrieved chunks with citations and tenant access in sanitized traces.
    • Design reversible quarantine and trust-regression tests without touching the live index.

    Inputs

    • Current corpus/index snapshot IDs, ingestion/embedding/reranker revisions and sanitized pipeline configuration
    • Source ownership, edit rights, import rules, document versions and approval metadata
    • Chunk manifests with stable document/chunk IDs, lineage and source access labels
    • Sanitized query/retrieval/citation traces, tenant/principal context and representative expected-source queries

    Outputs

    • Executive Summary
    • Scope
    • Assets / Trust Boundaries
    • Evidence Reviewed
    • Threat Model
    • Source Classification Ledger
    • Retrieval Attack Paths
    • Chunk And Citation Reconciliation
    • Findings
    • Critical Findings
    • Quick Wins
    • Prioritized Remediation Plan
    • Missing Evidence
    • Residual Risk
    • Verification Checklist
    • Unknowns
    • Next Validation Steps

    Permissions

    Read-only review of owner-supplied artifacts by default. Report file writes, shell/local fixture execution and official documentation lookup are optional, require available tools and their explicit scope. No live credentials are requested.

    Compatibility

    Standalone workflow, no Srednoff OS dependency. Codex evidence is synthetic entrypoint text only when recorded below. Claude Code, Cursor and actual buyer-runtime task delivery remain UNTESTED. Provider/platform-specific contracts must be refreshed for the installed version.

    Limitations

    This is an instruction/evidence workflow, not a hosted security service, persistent enforcement engine, production certification or business guarantee. The optional Python helper checks supplied claims, not underlying source artifacts. Agensi delivery is agent-only: do not save/export proprietary skill files to run local helpers; disclose NOT_EXECUTED where execution is unavailable.

    Validation Transparency

    Offline evidence-precheck status: PASS. Authored synthetic text-response cases: 16/16, status PASS. This is not independent security testing, complete artifact inspection or proof of deployment safety. Exact-text hashes and attempts are retained separately.

    License And Publisher

    Commercial License 1.0 covers original new expression; historical MIT and third-party rights remain unchanged. Mandatory marketplace conditions prevail. Publisher: Ivan Srednoff / Srednoff.Agency. No proprietary source resale or off-platform paid upsell is included.

    FAQ

    What is different from a generic security request? RAG data owner: document/source lineage ledger, ingestion authority classes, poisoned-metadata/citation mismatch cases and retrieval regression matrix. No generic prompt hardening, GEO ranking review, model retraining, vector-store scanner or runtime tool enforcement.

    Does it change production automatically? No. Audit/design is the default; changes, secret rotation, IAM, network probes and financial actions need separate explicit authorization.

    Are the examples real incidents? No. They are synthetic DEMO DATA.

    Is this a certification or guaranteed protection? No. Evidence gaps, residual risk and unexecuted tests remain explicit.

    Can I use project-specific reports for clients? Yes under the included license; proprietary workflow source is not a client deliverable.

    Version

    1.0.0

    Release Notes

    Initial scoped security workflow with domain decision rules, evidence requirements, synthetic examples, regression cases and explicit authorization/runtime limits. No production/customer outcome is claimed.

    This is an instruction/evidence workflow, not a hosted security service, persistent enforcement engine, production certification or business guarantee. The optional Python helper checks supplied claims, not underlying source artifacts. Agensi delivery is agent-only: do not save/export proprietary skill files to run local helpers; disclose NOT_EXECUTED where execution is unavailable.

    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

    Permissions

    Terminal / Shell
    Listed1 day ago

    What's inside

    Frequently Asked Questions

    Audit ingestion-to-citation trust using supplied artifacts and deliver a source classification ledger, retrieval attack-path review and regression plan. No live corpus scanner or poisoning detector is installed.

    This skill uses the open SKILL.md standard. Use it through an MCP-compatible agent connected to Agensi, and check the compatibility information on this page for confirmed agents and required capabilities.

    Connect your AI agent to Agensi through MCP, then ask it to use Srednoff RAG Poisoning & Retrieval Trust Auditor on your task. Your agent loads the latest approved version each time. Paid skills require a purchase or active subscription. Skill files are not available for download; your licence does not allow you to save, copy or export any part of a skill.

    New in Security