First-caller-wins in privileged setup
How an unclaimed initialization step can give permanent authority to the wrong caller. A pattern study with public precedent and defensive guidance.
FindingRead research
Independent research & developmentDesign · AI systems · Experiments
matching.work - Security research,systems & design
Sean Tehrani@back2matching
- Design
- Agents
- Systems
- Security
Security research
On-chain authority, agent trust, model supply chains.
AI systems
Agent tools, orchestration, and real-time infrastructure.
Design & experiments
Interfaces, generative art, and new protocols.
Security research
Selected public writeups. Methods, controls, and limits of the evidence are part of the work.
The MCP trust boundary includes the schema
A controlled study of why tool descriptions alone miss excessive agency, and what schema-aware checks can catch.
FindingRead research
What runs when you load a model?
Model-file execution paths and defensive scanner tests. Repackaging known malicious fixtures did not evade detection in the tested cases.
Negative resultRead research
Index14 writeups · newest first
- The Empty Boss ChairPattern class; live instances in coordinated disclosureOn-chainFinding
- The Effort Knob Is a Cost Knob0 / 420 gate evasions; effort moved cost, not securityAgents & MCPNegative result
- Which Zip Tools Are Actually Dangerous in WordPressPclZip escaped the root; 1,824 plugins checkedAppSecFinding
- Twin Fixtures, One VerdictDetection follows the payload, not the adapterModel supply chainMethod
- Chat-Template SSTIGGUF path matched standalone; two gadget families missedModel supply chainMethod
- MCP Excessive Agency in the Wild9 real servers, 5 detection classes, 0 exploitable bugsAgents & MCPNegative result
- MCP Audit Research Series, Batch 16 real servers, 2 new detection classesAgents & MCPMethod
- MCP Audit Research Series, Batch 2False positives are the next frontier, not recallAgents & MCPMethod
Profile · Stack
Security, systems & design · What I reach for
I research security boundaries in on-chain programs, AI agents, and the model supply chain.
My background is in design and full-stack development, so I approach systems through both their internals and the interfaces people use. This is a collection of published research, systems I've built, and experiments in design and technology.
Design & frontendWhere I started
- React
- Next.js
- SvelteKit
- GSAP
- Motion
- Design Systems
- Figma
- Typography
- Editorial Design
Security researchAuthority, boundaries, evidence
- Foundry
- Fuzzing
- Invariant Testing
- Slither
- Fork PoCs
- MCP Security
- Model Supply Chain
- Responsible Disclosure
AI & systemsTools with explicit scope
- Python
- TypeScript
- Multi-LLM
- Agent Orchestration
- NLP Pipelines
- Voice I/O
- Browser Automation
- Queues
- WebSocket
Backend & dataWhat keeps it running
- FastAPI
- Node
- Express
- Postgres
- Supabase
- Prisma
- Realtime
- Edge & Serverless
On-chain systemsContracts through interfaces
- Solidity
- Rust
- EVM & L2s
- Solana
- Account Abstraction
- EAS Attestations
- x402
- Foundry
How I Work
The bits that don't change
- i.
Start at the boundary.
Who can act, what they can reach, and where trust changes. Make those assumptions explicit before building on them.
- ii.
Test the claim.
Use reproducible fixtures, controls, and observable outcomes. Separate a detection result from a vulnerability, and publish the limits of the evidence.
- iii.
Design is part of the system.
Type, rhythm, and clear interactions matter. An interface should make a complex system easier to understand and use.
- iv.
Keep room for experiments.
Explore new protocols, generative art, and agent workflows. Label experiments honestly and carry useful lessons into the next build.
Selected Work
Project history · Solo · Systems & interfaces · Built & explored
Selected builds · AI systems, infrastructure, and generative artFour projects. From concept through implementation.
DOK
Deploy On Klik · Signal-to-contract pipeline
A pipeline connecting X signals to contract generation and Ethereum deployment. Built the ingestion, contract synthesis, deployment, and distribution workflow, with support for different launch patterns.
- Signal ingestion
- Contract generation
- Deployment workflow
TypeScript · Solidity · EVM · Next.js · CI/CD
Solo · Deployment infrastructure
PRISM
Agentic automated art framework
A generative art framework exploring cross-generational pattern evolution. A Python/Processing renderer produces animated loops; a gallery platform presents the work and connects it to a marketplace.
- Evolutionary pattern engine
- Animated renderer
- Gallery + marketplace
Python · Processing · Next.js · Supabase · Multi-LLM
Solo · Generative + gallery + market
cigoL
Reasoning agents · Prediction markets
Reasoning-agent framework for Polymarket's CLOB, built over months. Multi-LLM providers trade head to head with real-time benchmarking. Users deploy their own agents through Telegram. Gasless execution via EIP-7702 account abstraction, cross-chain deposit routing. A flow designed to handle gas for the user.
- Multi-LLM head to head
- EIP-7702 gasless AA
- Cross-chain deposits
Python · Next.js · Polymarket · wagmi · Supabase
Solo · cigol.ai
FlockRun
Hierarchical agent orchestrator
An internal multi-agent orchestrator with a coordinating agent and specialised workers for research, development, and operations. Includes Chrome DevTools Protocol browser tooling and a live WebSocket monitoring dashboard.
- Task delegation
- CDP browser tooling
- Live monitoring dashboard
TypeScript · Node · SQLite · BullMQ · CDP · WS
Solo · Internal tool
Open source · PyPI
TurboQuant & tools
Independent TurboQuant KV-cache compression and the benchmarking tools built around it.
Read the benchmarks- turboquantKV-cache compression for Hugging Face models
pip install turboquant - turboquant-vectorsEmbedding compression
pip install turboquant-vectors - kvcache-benchOllama KV-cache benchmarks
pip install kvcache-bench - quantsim-benchOllama quant-level benchmarks
pip install quantsim-bench
Experiments
New protocols · Complex mechanics · Built & explored
Snakey
x402 multiplayer game · Archived reference
An archived multiplayer game and SDK exploring x402 payments between agents and humans. The reference implementation uses Base Sepolia testnet; it was not launched on mainnet.
- x402 payment experiments
- Published npm SDK
- Testnet reference
JavaScript · Express · SvelteKit · Base Sepolia
Solo · Base Sepolia testnet
XFRAME
Trading terminal · Live streams
Full trading terminal. TradingView charts, on-chain whale tracking, social sentiment aggregation, KOL activity monitoring. Live streaming market data over WebSocket, gamified leaderboard, mobile-optimised.
- WS market stream
- Whale + sentiment fusion
- Gamified leaderboard
Next.js · TypeScript · TradingView · Privy · Solana
Solo
Medari
Multi-modal health AI
Multi-modal health platform covering consultations, fitness, nutrition, mental health. Voice pipeline with STT/TTS, gamification engine with tiered achievements, hybrid fiat + crypto payment rail.
- Voice I/O pipeline
- Gamification engine
- Hybrid payment rail
Next.js · React · Multi-LLM · Supabase · Prisma
Solo
SIGINT
Real-time crypto intelligence
Ingests live Telegram feeds across chains. Runs NLP entity recognition and signal extraction on every message, then correlates caller reliability against actual on-chain price action. Tiered alerts, scheduled AI intel reports. A trust engine built on price reality.
- NLP signal extraction
- Trust-scoring engine
- Automated intel reports
Python · Telethon · PostgreSQL · Multi-LLM
Solo
Clawsight
Agent analytics on Base
Agent analytics and ad marketplace on Base, with on-chain reputation via EAS attestations and verified contract deployment. An experiment in connecting agent activity to attestations and a public interface.
- EAS attestations
- Verified contracts
- Agent-native analytics
Base L2 · Solidity · EAS · TypeScript
Solo · On-chain reputation
Reach
Remote · async.