Security
Intelligence.

Stellar's Soroban smart contract platform encodes all function calls as XDR binary — a format hardware wallets cannot decode or display. Where Ethereum has ERC-7730 clear signing descriptors, Soroban has no equivalent. This paper maps the attack surface, documents the gap, and specifies exactly what a new SEP for Soroban clear signing would define — the descriptor format, registry model, trust architecture, and wallet behaviour standard the ecosystem needs to close it.

In February 2025, three Bybit co-signers approved a Safe transaction their hardware wallets displayed as unreadable hexadecimal data. Attackers had replaced the multisig implementation contract. The resulting theft — $1.5 billion — is the largest single cryptocurrency theft in history. This paper examines the structural blind signing vulnerability, maps the loss record, and details the ERC-7730 clear signing descriptor standard and Patrick Collins' ERC-8213 bytes-level fallback mechanism that the Ethereum community is deploying to close it.

The security industry built its tooling for static, deterministic smart contracts. Those contracts are now embedded in ecosystems with autonomous AI agents, LLM-driven governance, multi-chain bridges, and small teams vulnerable to fabricated video calls. Only two of the five active attack layers fall within traditional audit scope. This paper presents a unified five-layer threat model and maps the security disciplines required to defend each layer.

In November 2024, an AI agent holding a $47,000 prize pool was publicly broken in 481 prompts. The winning attack did not crack cryptography — it redefined function semantics inside the agent's reasoning context until it approved a transfer it was designed to refuse. This paper maps the adversarial manipulation threat to production AI trading agents, analyses the Freysa proof-of-concept, and provides a security architecture for teams deploying autonomous on-chain agents.

Ethereum L1 processes 15 transactions per second. A real-time AI inference pipeline may handle 50,000 requests per second. The 1,600x speed gap between blockchain consensus and AI compute cannot be closed — it can only be bridged with off-chain integration layers. Those layers are where $1.3 billion in cross-chain bridge losses have been recorded, and where rushed AI integrations leave open APIs and race conditions that are trivially exploitable.

A finance employee wired $25 million after a deepfake video call featuring a convincing AI recreation of the company CFO. No password was stolen. No exploit was run. The attack was entirely social. This paper examines how deepfake technology specifically threatens Web3 key management, why AI-assisted development teams are at heightened risk, and what procedural and technical controls can stop it.

Every ECDSA private key in existence is theoretically vulnerable to a sufficiently powerful quantum computer running Shor's algorithm. That computer does not yet exist — but the adversaries planning to use it are already collecting encrypted data today. This paper examines the timeline, the exposure surface, and the migration path for Web3 protocols that want to survive the transition.

As protocols integrate AI oracles, LLM-powered governance, and on-chain model inference, they inherit an entirely new class of vulnerabilities that traditional audit methodologies are not equipped to catch. This paper maps the emerging attack surface and proposes a framework for evaluating AI-integrated smart contract systems.

Security audits are necessary but not sufficient. A manual review finds bugs that exist in the code as written — formal verification proves properties that must hold for all possible inputs across all possible states. This paper makes the case that the most critical DeFi protocols should treat formal verification not as a premium add-on, but as a baseline requirement.