TechFlow reported on June 11 that Agora, an automated testing framework proposed by 0G Labs together with research teams from the National University of Singapore, Peking University, Beijing University of Posts and Telecommunications and other institutions, has been accepted by ICML 2026. The framework is designed for automated vulnerability detection in production-grade consensus protocols and, according to the report, is the first to deeply combine distributed systems domain knowledge with a Multi-Agent collaborative architecture for this purpose.
Agora Targets Deep Bugs in Production Consensus Protocols
According to the paper, Agora has discovered 15 previously unknown deep logical vulnerabilities, described as Deep Bugs, in mainstream consensus protocols including Raft, EPaxos, HotStuff and BullShark. The issues involve key security areas such as execution divergence, monotonicity violations, topology defects and signature verification. These categories relate directly to the correctness and safety of consensus mechanisms operating under complex distributed conditions.
The research results also state that, under the same testing scenarios, mainstream large language models such as GPT-5.2 and Claude 4.5 did not find the related protocol-level vulnerabilities. The comparison presented in the paper places emphasis on Agora’s combination of distributed systems knowledge and structured testing procedures, rather than relying only on general-purpose model analysis.
Hypothesis-Driven Testing and Multi-Agent Collaboration
Agora uses Hypothesis-Driven Testing together with a Multi-Agent collaboration mechanism. The framework automatically generates attack scenarios, executes tests and dynamically revises its testing process, enabling deeper security analysis of complex distributed systems. In this workflow, testing hypotheses guide the exploration of protocol behavior, while repeated execution and adjustment help expose logical defects that are difficult to identify through ordinary testing.
Beyond consensus protocols, the framework is also described as extensible to database concurrency control, operating system kernels and Web3 smart contract auditing. Its acceptance by ICML 2026 places the consensus protocol debugging agent framework, jointly developed by 0G Labs, the National University of Singapore, Peking University, Beijing University of Posts and Telecommunications and other research teams, within the paper track of a leading machine learning conference.