Network
Bradbury testnet
Living security advisory · v2026.05.13
Eight failure modes of Intelligent Contracts, proven on-chain.
A field reference for engineers writing GenLayer Intelligent Contracts. Each module pairs a vulnerable contract with its patched twin and links the real Bradbury testnet transactions that demonstrate the bug failing and the fix succeeding. No theoretical exploits — every claim resolves to a finalized receipt.
Incidents · 01 – 08
8 modules · 16 questions
M.01
URL Rot
The Disappearing Oracle
ATTACK
Hardcode a URL. Site goes down or changes anti-bot policy. Contract fails forever.
Read incidentFIX
Domain whitelisting + fallback list. Maintain multiple oracle endpoints.
Read incidentM.02
Prompt Injection
The Trojan Prompt
ATTACK
User input contains hidden instructions that hijack the LLM output.
Read incidentFIX
Input sanitization + greyboxing pre-filters. Never pass raw user input directly to the LLM prompt.
Read incidentM.03
API Key Leakage
The Transparent Secret
ATTACK
Hardcode OPENAI_API_KEY in contract. Validators can read it.
Read incidentFIX
Never put secrets in contracts. Use proxy patterns or off-chain pre-processing.
Read incidentM.04
Wrong Equivalence
The Precision Trap
ATTACK
Use strict_eq for live ETH price. Validators get slightly different numbers. Consensus fails.
Read incidentFIX
Use custom equivalence with tolerance ranges for volatile data.
Read incidentM.05
Missing Access Control
The Open Door
ATTACK
@gl.public.write method has no owner check. Anyone can drain state.
Read incidentFIX
require_sender(self._owner) or role-based access control.
Read incidentM.06
Cloudflare Block
The Invisible Wall
ATTACK
Target URL is behind Cloudflare. web.render() returns challenge page, not content.
Read incidentFIX
Pre-verify URLs with Intelligent Crawler. Maintain on-chain URL health registry.
Read incidentM.07
Biased Prompt
The Loaded Question
ATTACK
Author embeds a hardcoded answer in the system prompt, so the LLM ignores the actual review content.
Read incidentFIX
Deterministic lexicon pre-classifier gates the LLM. When the lexicon is decisive, no LLM call is made and prompt bias has no path to the verdict.
Read incidentM.08
URL Spoofing
The Fake Source
ATTACK
Malicious actor spins up a fake news site clone to feed false data.
Read incidentFIX
Domain whitelisting as mutable state, not static code. Community governance for domain additions.
Read incidentHow to read this playbook
01 · Read
Open an incident
Each module page describes the failure mode in plain language and shows the vulnerable code beside the patched code.
02 · Verify
Follow the receipts
Two transaction-link buttons take you to Bradbury Explorer — the failing call and the succeeding call, each with execution result inline.
03 · Check
Answer the quiz
Two questions per incident lock once committed. The combined quiz aggregates all 16 for a final score.