Phala Network — agentic threat model

Not certain from the listing — Phala acts as an execution infrastructure rather than providing a specific foundation model. Standard LLM threats like prompt injection or alignment issues would depend entirely on the developer's choice of model deployed within the TEE.

Data operations leverage Trusted Execution Environments (TEEs) to guarantee confidentiality and prevent unauthorized data exfiltration or state tampering during execution. However, data poisoning of the smart contracts or external feeds (oracles) feeding the agents remains a critical threat.

The framework facilitates AI Agent integration with smart contracts. The primary threat here is insecure tool integration, where an agent might be manipulated into executing unauthorized or malicious smart contract transactions (e.g., unauthorized token transfers).

Deployment relies on decentralized cloud nodes running TEEs. While TEEs mitigate host compromise and lateral movement, they are susceptible to hardware-level side-channel attacks or zero-day exploits in the enclave software.

Not certain from the listing — While blockchain transactions are transparently recorded, the internal reasoning and execution logs of the AI agents running inside the confidential TEEs may suffer from severe observability and auditing blind spots.

Security is enforced cryptographically via blockchain consensus, staking, and hardware attestation. However, compliance with traditional financial regulations (KYC/AML) is highly complex due to the decentralized and privacy-preserving nature of the network.

The ecosystem involves multi-agent and smart contract interactions. Rogue or compromised agents could exploit trust relationships with other agents or smart contracts, potentially triggering cascading financial failures across the decentralized network.