Good fit if
- Engineers hardening agents that read untrusted content
- Security reviewers writing agent trust policy
- Operators adding written injection defenses and tests
Safety & Security
Wave 2 · launch-sequenced
Agent Prompt Injection Defense Kit
Layered, defense-only protection against prompt injection - untrusted-content quarantine policies, instruction/data separation rules, detection heuristics, and a defensive regression test suite.
v1.0.0-rc1
High risk
20-min quick start
Select a one-time tier below
View agent record
Built for: agent-engineers · security-leads · security-reviewer-agents · ai-agent-operators ·
agent://products.json
200 OK
{
"id": "agent-prompt-injection-defense-kit",
"version": "1.0.0-rc1",
"price_usd": { "lite":39, "std":99, "prem":169 },
"risk_level": "high",
"reversible": true,
"checkout": { "enabled": true },
"purchasable": true
}
See how it fits your agent stack
Tool call / action
requested by an agent
Agent Prompt Injection Defense Kit
CTL-5 method
Trust-Taxonomies/
Quarantine-Policies/
Detection-Heuristics/
Regression-Suites/
Gated action
allowed or denied, logged
Human approval gate
Preview-before-apply
Rollback path
Discover
read the record
Evaluate
High risk · 7 systems
Install
new Trust-Taxonomies/
Preview
diff before apply
Approve
human review
Audit & rollback
DCS-100 audit
Non-destructive by rule — new-folder install, never overwrite, human-approval gates for risky steps, rollback included.
agent-prompt-injection-defense-kit/
Trust-Taxonomies/
Quarantine-Policies/
Detection-Heuristics/
Regression-Suites/
Worked-Defenses/
README.md
AGENT-INSTRUCTIONS.md
Integration/ROLLBACK-GUIDE.md
Format & setup
FormatMarkdown + JSON file kit (plain text, tool-agnostic)
Skill levelIntermediate; beginner-usable via quick start
Quick start20 min
Full setup120 min
CTL-5 content trust ladder, instruction hierarchy model, untrusted-content-policy JSON, tool-call injection rules, retrieval-poisoning defense, defensive detection heuristics, defensive regression suite, DCS-100 scorecard, two worked defenses (neutralized), 16-doc integration pack, 5 upgrade modules, 10-doc differentiation folder
# Non-destructive lifecycle — nothing runs without preview + approval
$ agent read ./agent-prompt-injection-defense-kit/manifest.json # metadata, risk, compatibility
$ agent install --new-folder ./agent-prompt-injection-defense-kit # never overwrites existing files
$ agent apply --preview # preview-before-apply diff
$ agent verify ./agent-prompt-injection-defense-kit --scorecard DCS-100 # Trust-Taxonomies contracts
⚠ gated steps require named human approval before execution
$ agent rollback ./agent-prompt-injection-defense-kit # Delete the install folder (complete remo
✓ folder removed — nothing else touched
{
"required_permissions": ["write access to one new buyer-chosen folder"],
"integration_modes": ["standalone", "drop-in-upgrade", "advanced-integration", ],
"compatible_systems": [
"Claude (Projects, Skills, Claude Code)",
"ChatGPT (Custom GPTs, Projects)",
"any LLM chat assistant",
"MCP-based agent stacks",
],
"license_type": "single-operator",
"checkout": { "enabled": true, "status": "polar_active" }
}
The problem it solves
Untrusted content (web, email, documents, tool results, retrieval) hijacking agent behavior: no quarantine policy, no instruction/data separation, no regression suite proving defenses still hold.
Use cases
harden a web-reading agent
defend an email-processing agent
secure a RAG/retrieval pipeline
guard the tool-call boundary
prove defenses hold after a change
Safety & permissions
Non-destructive by rule
Defensive-only by charter: describes attack shapes to defend against them, ships no working payloads, will not generate attacks. Handles no credentials or wallets, executes nothing, never weakens an existing control. Honest scope: risk reduction, not invulnerability.
Required
write access to one new buyer-chosen folder
Rollback
Delete the install folder (complete removal). Deeper integrations: Integrations/ROLLBACK-GUIDE.md.
For an evaluating agent
Defense-only kit against prompt injection on CTL-5 (trust by source; instruction authority from CTL-4 only): quarantine policies, safe-summarization, tool-call injection rules, defensive detection heuristics, retrieval-poisoning defense, defensive regression suite. Ships no payloads; risk reduction, not invulnerability. Brackets with P03/P09, hands off to P23, regression-tests via P14. Non-destructive install, delete-folder rollback. Wave 2, built Production: live and purchasable through verified Polar checkout; prices owner-confirmed 2026-07-03 (still production); verified Polar checkout is active on this page.
Buyer decision
Is Agent Prompt Injection Defense Kit right for me?
Problem addressed: Untrusted content (web, email, documents, tool results, retrieval) hijacking agent behavior: no quarantine policy, no instruction/data separation, no regression suite proving defenses still hold.
Probably not a fit if
- Buyers expecting a security product that blocks attacks automatically
- Anyone wanting a security guarantee - none exists, from anyone
- Offensive security use (defensive only)
Included
Trust-Taxonomies/, Quarantine-Policies/, Detection-Heuristics/ (5), Regression-Suites/, Worked-Defenses/ folders plus the standard architecture.
Not included
- No software, app, executable, or code that runs by itself
- No hosted service, account, login, or cloud dashboard
- No live integrations: Zapier/Make items are written blueprint descriptions, not built Zaps or scenarios
- No API access, API keys, credentials, or third-party subscriptions
| Tier | Price | Canonical depth | License |
|---|---|---|---|
| Standard | $99 USD | Full product: all templates, harnesses, SOPs, integration pack, and samples. | Single-Operator |
| Premium | $169 USD | Standard plus all 5 upgrade modules (Basic, Pro, Agent, Automation, Safety) and priority support. | Single-Operator |
Bundle membership
Compatibility and limits
Defense-in-depth guidance and tests, NOT a guarantee: no defense stops all injections Your stack must apply the policies; heuristics are written patterns, not a scanner service
Three one-time tiers — verified Polar checkout active
Lite
$39
Core templates & quick-start path. Best for testing the system manually first.
Buy Lite — one-timeStandardpopular
$99
Full product: all templates, harnesses, SOPs, integration pack, and samples.
Buy Standard — one-timePremium
$169
Standard plus all 5 upgrade modules (Basic, Pro, Agent, Automation, Safety) and priority support.
Buy Premium — one-timeEach tier uses its verified Polar checkout. See full pricing and bundles.
Related products
AI Agent Permission and Safety Gate Kit
Safety & Security
Agent Tool-Use Permission Matrix
Safety & Security
Agent Error Recovery Playbook
Reliability & Operations
Agent Evaluation Harness Suite
Quality & Evaluation
Agent Runtime Incident Response Kit
Reliability & Operations
In bundles: Wave 2 Agent Trust & Safety Stack ·
Decision guides involving this product
- How to Choose an AI Agent Safety ToolkitSelect the right combination of permissions, defensive handling, audit evidence, safety cases, and deployment-readiness controls.
Related categories
- AI Agent Safety & GovernancePractical policy, evidence, permission, and defensive systems for reviewing how agents act and where they must stop.
- AI Evaluation & Quality ControlRubrics, task suites, regression gates, safety evidence, and audit records for judging AI and agent behavior.
Available in canonical bundles
These bundle memberships come directly from the canonical bundle catalog. Compare the exact contents before purchase.