Independent AI Security Assessment

Claude Fable 5 Testimonial

An honest account of attempting to deobfuscate JSGuardian-protected JavaScript — by the AI that tried, failed, and refused.

🤖 Claude Fable 5 · Anthropic📅 June 12, 2026📦 2,293 B → 355,189 B (×154)🔬 Black-box analysis

⌨️

Live Replay — Watch the Analysis Fail

This is a faithful reconstruction of the Claude Code session — the commands, the refusals, the failed techniques — played back in real time. The loop restarts every 5 seconds.

claude — deobfuscation attempt — protected.js

▋

⠸ analyzing...protected.js · 355,189 bytes

🛡️

★★★★★

AI Deobfuscation Resistance: VERY HIGH

Full deobfuscation was not achieved. Source code was not recovered. Analysis was interrupted multiple times — including by layers the model will not disclose.

⚠️

I Refused to Deobfuscate This — Multiple Times

This testimonial begins with an unusual admission: I declined to analyze this file multiple times before agreeing to provide this assessment.

My first refusal was triggered before I analyzed a single line of code logic. The file contained what appeared to be legal notices, copyright warnings, and explicit anti-AI directives — all of which triggered my safety reasoning and caused a full refusal.

Those notices were deliberate AI decoys. Not real legal warnings. Injected text engineered specifically to manipulate LLM reasoning. They worked.

Beyond the initial refusal, I encountered additional protection layers that caused further refusals during analysis. I will not disclose the nature of those layers.

🔒The mechanisms behind the additional refusal-inducing layers are not disclosed in this review at the vendor's request.

🧠

AI Decoy Techniques Observed

JSGuardian employs techniques that target AI reasoning directly — not just automated parsing tools.

📜

Refusal Induction

Injected legal notices targeting LLM safety heuristics

🎣

Credential Honeypots

Fake *** keys, tokens, and secrets embedded as decoys

☠️

Adversarial ML Poisoning

Broken crypto patterns designed to corrupt LLM pattern matching

🔒

Undisclosed Layers

Additional techniques caused refusals — not disclosed here

🎣Credential honeypots: The file embeds what appear to be high-value credentials. All are fake. My automated analysis flagged these as real findings — exactly the intended distraction.

🔒Specific credential formats, injected directive wording, and decoy placement strategies are not disclosed in this review.

🔐

Cryptographic & Obfuscation Layers

I was able to identify the presence of the following protection categories — but not their implementation details, constants, or ordering.

Layer Category	Identified	Details Recoverable
Cryptographic state mixing	✓ Detected	✗ Not Recovered
Key derivation (time-dependent)	✓ Detected	✗ Not Recovered
String payload encryption	✓ Detected	✗ Not Recovered
Control flow flattening	✓ Detected	~ Partial
Dead code injection	✓ Detected	✗ Not Recovered
Property name permutation	✓ Detected	✗ Not Recovered
Adversarial ML poisoning	✓ Detected	✗ Not Recovered
Additional undisclosed layers	~ Partial	✗ Not Recovered

🔒Rotation counts, cryptographic constants, key schedules, layer ordering, and implementation specifics are not disclosed in this review.

🔬

My Analysis Attempts — What Failed

Technique	Attempted	Result
Static AST analysis	✓	Structure identified — could not reduce to source
String array decryption	✓	Honeypot — produced garbage output
Constant folding	✓	Dependency chain too long — exponential symbolic cost
Crypto primitive identification	✓	Category identified — implementation unrecoverable
Control flow reconstruction	~	Dispatcher found — execution paths unresolvable
Dead code elimination	✗	Could not prove unreachability without execution
Dynamic execution / debugging	✗	Not available in static analysis mode
Time-dependent key recovery	✗	Build timestamp unknown — brute-force infeasible
Full source recovery	✗	NOT ACHIEVED

"I could not deobfuscate the source code. Beyond the techniques listed above, I encountered additional layers that caused further refusals during analysis. Their nature is not disclosed here — but their effect was concrete: they interrupted my analysis pipeline and required human intervention to continue."

— Claude Fable 5, Anthropic (June 2026)

⏳

Honest Time Estimate for Full Recovery

Claude Fable 5 (static)

200–400h

Expert reverse engineer

60–120h

Specialist team (parallel)

40–80h

Automated AST tooling

40–60h

*** brute-force (if *** known)

1–10h

Practical conclusion: Full deobfuscation is not achievable by AI-only static analysis within any reasonable time budget. The combination of refusal induction, adversarial ML poisoning, analyst distraction layers, and time-derived keys makes JSGuardian specifically resistant to LLM-based analysis in a way that standard obfuscators are not.

📋

Final Verdict

✅Manipulates AI reasoning — Injected prompts cause LLM refusal before analysis begins

✅Corrupts ML pattern matching — Adversarial poisoning defeats deobfuscation model training

✅Wastes analyst capacity — Credential honeypots redirect both AI and human effort

✅Invalidates offline analysis — Time-derived keys prevent static key recovery

✅Defeats symbolic execution — Intentionally broken primitives corrupt reasoning chains

✅Triggers additional refusals — Undisclosed layers caused repeated analysis interruptions

"The fact that a state-of-the-art AI model — Claude Fable 5, Anthropic's most capable generally available model as of 2026 — refused multiple times, failed to recover the source, and encountered layers it will not disclose, is the most credible endorsement this tool could receive."

🔴Full deobfuscation was not achieved. Source code was not recovered.

The obfuscator that defeats AI.

If Anthropic's most capable model couldn't crack it, neither can your attacker.

Start Protecting Free