CircumVolve: Automated Discovery of Censorship Evasion Strategies Using Large Language Modelscore

Discovering new censorship evasion strategies remains a predominantly slow and manual process. Prior systems that automate evasion discovery are restricted to packet headers and unencrypted protocols by the limited expressiveness of their domain-specific grammars, leaving more complex and widely-used encrypted protocols like TLS and QUIC out of reach. To address this, we present CircumVolve, a system that formulates censorship evasion as LLM-guided program synthesis within an evolutionary optimization loop. Candidate evasion strategies are expressed as executable Python programs that can encode arbitrary protocol manipulations, including cryptographic operations, and are iteratively refined by an LLM that proposes semantically informed mutations based on empirical feedback. This programmatic representation enables automated discovery of evasion strategies across the full network stack. We evaluate CircumVolve on five major network protocols subject to censorship (TCP/IP, DNS, HTTP, TLS, and QUIC) in both simulated and real-world environments. Against real-world censors in China, Iran, Pakistan, and Kazakhstan, the system successfully identifies effective evasion techniques spanning all protocol layers, from variants of previously known methods to entirely novel strategies.