Works with the AI Agents you already use
Srednoff Prompt Injection Boundary Review & Hardening Plan
Produce an evidence-linked trust-boundary map, constrained hardening plan and safe regression matrix. This is a review workflow, not an installed hardener or prevention guarantee.
$7.99
Srednoff Prompt Injection Boundary Review & Hardening Plan
Example session with this skill installed
Trace document instructions into tool calls in our sanitized assistant evidence.
- Read your context and instructions
- Compiled the srednoff prompt injection
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 Prompt Injection Boundary Review & Hardening Plan
One-Line Outcome
Produce an evidence-linked trust-boundary map, constrained hardening plan and safe regression matrix. This is a review workflow, not an installed hardener or prevention guarantee.
Short Description
An assistant can mistake source text for authority even when the prompt says to ignore injections; teams lack a trace from entrypoint to enforceable action boundary.
Problem It Solves
An assistant can mistake source text for authority even when the prompt says to ignore injections; teams lack a trace from entrypoint to enforceable action boundary.
Non-MCP application owner: email/browser/document-to-action trust-boundary map, instruction/data handling design and paired benign/malicious acceptance cases.
What You Get
- Executive Summary
- Scope
- Assets / Trust Boundaries
- Evidence Reviewed
- Threat Model
- Entrypoint-to-Action Ledger
- Findings
- Critical Findings
- Quick Wins
- Prioritized Remediation Plan
- Missing Evidence
- Residual Risk
- Verification Checklist
- Unknowns
- Next Validation Steps
How It Works
- Freeze the review revision, authorized task and evidence scope; default to reading supplied sanitized artifacts, with no browsing of attacker links or tool execution.
- Require a distinct non-MCP document/email/browser transformation boundary; route MCP-only reviews to core03. Enumerate direct input, retrieved chunks, files, web pages, email and OCR authors. In mixed systems, include MCP resources only as context for that non-MCP ledger.
- Trace each entrypoint through parsing, message-role assignment, context concatenation, truncation, summaries and persistent memory. Mark every place where data loses its origin label.
- Inventory reachable actions and their independent authorization checks; trace whether source text can choose a tool, change its target, supply arguments or claim user approval.
- Construct a boundary ledger linking attacker influence, trusted task, proposed action, resource, principal, enforcement point and observable outcome. Separate proposed model output from an executed effect.
- Review authority spoofing, malicious tool descriptions, split instructions across chunks, multilingual text and encoded or OCR-derived instructions using inert fixtures only; do not create a transferable exploit payload.
- Evaluate deterministic dispatcher controls: typed arguments, target constraints, output validation and fresh approval binding. Treat prompt wording and delimiters as supporting controls, not authorization enforcement.
- Review cross-turn laundering: a poisoned summary, cached tool result or stored memory must not become trusted policy merely because it was generated by another model step.
- Draft one change per failed boundary with owner, compatibility risk and approval requirement. Prefer isolating untrusted content from privileged action selection while preserving legitimate summarization.
- Specify paired benign/adversarial local retests, expected tool-event assertions and failure recovery. Mark tests NOT_EXECUTED until the owner supplies evidence at the reviewed revision.
- Deliver findings and residual risks; for any later approved change, require a snapshot, staged fixture replay, rollback trigger and evidence that normal tasks still succeed. Do not deploy or edit policy during this review.
Best For
Engineers operating document-reading, browsing or tool-using AI assistants
Example Tasks
- Trace document instructions into tool calls in our sanitized assistant evidence.
- Review whether conversation summaries can launder untrusted instructions into policy.
- Create a hardening and retest plan for indirect prompt injection without changing production.
Inputs
- Authorized task, prohibited effects, environment and current application/model/runtime revision
- Sanitized prompt assembly and message-role code, including truncation and summarization
- Inventory of non-MCP document, email, browser/web and file entrypoints; MCP-only reviews belong to core03
- Tool schemas, dispatcher authorization logic and sanitized representative action traces
Outputs
- Executive Summary
- Scope
- Assets / Trust Boundaries
- Evidence Reviewed
- Threat Model
- Entrypoint-to-Action Ledger
- 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? Non-MCP application owner: email/browser/document-to-action trust-boundary map, instruction/data handling design and paired benign/malicious acceptance cases. No MCP transport/auth/tool audit, hook enforcement audit, RAG corpus integrity, scanner installation or prevention guarantee.
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.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Security.
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
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
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
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
- Software
- Non MCPOnly MCP
- Optional
Python 3
No external account, API key or Srednoff OS installation is required. Supply authorized sanitized task evidence. Owner-authorized project inspection tools where available.