JSO AI is the AI feature surface JSO is shipping in three phases over the next nine months. This page is the customer-facing overview: what's coming, why now, how the tiers work, what the position is. For the engineering strategy behind it, see AI_PLAN.md; for the threat-model framing it responds to, see the CASCADE blog post.
New? The 5-minute quick start walks you from zero to your first AI call with curl + Python copy-paste fragments. Try it without auth: PresetAssistantPreview (describe app, get jso.config.json), CompatCheckPreview (paste JS, see compatibility findings), ExplainErrorPreview (paste a runtime error, diagnose the transform). All client-side rule-based prototypes that mirror the Phase 1 UX. The LLM-backed versions ship 2026-Q3 behind the same response envelope; your client code does not change.
What you're going to be able to do
Preset assistant. Type "Protect my React SaaS app with balanced performance" in plain English. JSO AI writes the matching jso.config.json. Edit before applying; we never auto-commit a config without confirmation.
Compatibility checker. Paste a bundle or point at a repo before you protect. The checker flags risky patterns (eval, dynamic require shape, framework runtime files that historically break under Maximum mode) and suggests exclusions. Catches the things that used to surface as a broken build in production.
Resistance Score. Every protected build gets scored against a CASCADE-style adversarial probe. The number isn't marketing — it's a concrete recovery percentage from an LLM attacker run over your protected output. If the score drops, the dashboard explains which option to flip to bring it back up. Phase 2.
Selective-obfuscation suggester. Most code in a typical bundle doesn't need Maximum mode — vendor files, framework runtime, polyfills. The handful of functions that do (license validation, anti-tamper, payment guards, key derivation) deserve VM virtualization, which costs runtime. AI identifies which is which and proposes a targeted protection profile so you don't pay the Maximum-mode cost on code that doesn't need it. Phase 2.
AI Deobfuscation Benchmark. AI Enterprise customers get a per-build report: "on release rel-abcdef123, CASCADE recovered 12% of original function names and 0% of string literals. JSIMPLIFIER recovered 8% and 0%. Both deobfuscators rated control-flow recovery as failed." Honest customer-specific numbers your security team can cite in a vendor review. Phase 3.
Config-help chat. Paste an error, a stack trace, a snippet of broken bundle output. JSO AI explains which transform plausibly caused the issue and which option to flip. Backed by the JSO option-table reference so the suggestions are grounded.
Why it's a separate subscription
- Customers who bought JSO for the obfuscation engine shouldn't pay for AI features they don't want.
- Customers who want AI should pay proportionally to how much they use it — not bundled into an obfuscation tier that mixes the two.
- Separability lets us iterate on the AI pricing without re-pricing the core product.
You can have any combination — Free obfuscation + AI Basic, Corporate obfuscation + AI Corporate, etc. The tiers are independent. Corporate obfuscation customers get AI Basic free for the first month after upgrade as a thank-you.
How usage is metered
Two units, both displayed live on the dashboard, both hard-capped by default:
- AI actions — the UX events. One preset suggestion = 1 action. One compatibility analysis = 1 action. One full Resistance Score audit = 5 actions. One minute of back-and-forth chat = 1 action.
- Protection token pool — tokens consumed by at-protect-time LLM work (variant generation, selective obfuscation suggestions). Counted because input size matters; a 10-line file isn't the same cost as a 50K-line bundle.
Hard quota; no surprise overages. When you exhaust either pool the AI features degrade to "explain the action and link to the upgrade path" mode. They don't silently bill more. If you want overage, it'll be an opt-in toggle (not in the first ship); for now, monthly buckets work the way the dashboard says they do.
Where the data goes
- Customer code submitted to AI features is processed in memory under the same no-persistence policy as the obfuscation pipeline. Snippets are not stored after the request completes.
- Underlying LLM provider: Claude (Anthropic) as primary for code understanding; OpenAI GPT-class as fallback. Both are configured with the providers' no-training-on-customer-data settings.
- The exception: the Resistance Score output itself is stored alongside your protection reports under the same retention policy — you need to be able to look up "what was the score on rel-abcdef123" weeks later.
- Every AI feature surfaces a clear "this snippet will be sent to {provider} for {task} and not stored" notice before submission. There's an account-level Disable AI toggle for organizations that want to opt out entirely.
Roadmap
| Phase | Ship | Features | Tier |
| Phase 1 — UX foundations |
2026-Q3 |
Preset assistant, compatibility checker, config-help chat. Subscription + metering plumbing. |
AI Basic+ |
| Phase 2 — Differentiation |
2026-Q4 |
Resistance Score, selective-obfuscation suggester, AI variant generation (light), pre-protection malware scanner. |
AI Corporate+ |
| Phase 3 — Moat |
2027-Q1 |
AI Deobfuscation Benchmark on customer code, continuous anti-LLM evolution dashboard, full AI variant generation, named SRE contact + SLA. |
AI Enterprise |
Customer demand on the waitlist directly weights this. Enterprise pilot requests get bumped to the front; AI Basic waitlist volume drives Phase 1 scope decisions.
The position
JSO is not selling "AI-proof" obfuscation. That claim doesn't survive any competitor blog post and we're not going to make it. JSO is selling AI-resistant, AI-aware, threat-model-honest protection — backed by a published response to CASCADE, a public roadmap that names the attack class each phase addresses, and (from Phase 2 onward) a customer-verifiable Resistance Score number on every protected build.
The Resistance Score is the artifact that makes the claim auditable. The Deobfuscation Benchmark is the artifact that makes it customer-specific. Neither replaces the existing protection — they describe how well it's holding up against the real-2026 attacker.
Join the waitlist
The waitlist forms below feed straight into our development priority queue. We use them to size Phase 1 scope and to schedule Phase 3 Enterprise pilots.