Research journal · Announcement

Introducing Wake 5

Wake 5 fuzzes Solidity from one prompt: an agent writes the fuzz test, an auditor reviews it in Python, and an EVM inside the Python process runs it as fast as Foundry.

research/introducing-wake-5 figure / cover
Wake 5 card: the Wake mark, the words Wake 5, the command pip3 install eth-wake, github.com/ack3-ai/wake, developed by ack3, arXiv 2611.09307.
The closing card of the Wake 5 launch clip.

Wake 5, the open-source Python framework ack3’s auditors use to test, fuzz and statically analyse Solidity, is available today:

pip3 install eth-wake

The release changes who writes the fuzz test. In manually guided fuzzing, auditors write the harness: the flows, a Python model of the contract’s state, and the invariants that must hold. In Wake 5 a coding agent writes that harness from one prompt, the auditor reviews it as ordinary Python, and Wake runs it at native speed. The prompt is in the Wake README.

What is new

  • An EVM inside the Python process. The default engine is revm, bound to Python. No development node to start; live chains fork through chain.connect(fork=...), and several chains run in one process for cross-chain tests.
  • An MCP server for agents. wake mcp gives an agent the static facts the language server has: contracts, definitions and references, inheritance, storage layout, state changes, selectors, external calls.
  • Multi-hour campaigns stay memory-stable. Chain history is bounded to the newest 256 blocks by default.
  • Generated Python types for every contract, so tests and agents get autocompletion and type checking against the Solidity code. Types generated by Wake 4 need regenerating; the migration guide lists every breaking change.

Measured in the preprint

The preprint Wake 5: A Python Framework for Solidity Testing, Fuzzing, and Static Analysis by Josef Gattermayer and Michal Převrátil describes the design and measures it on the five L32 mazes of the Daedaluzz benchmark: 153 reachable assertions behind guards built from constants.

Two bar panels from the Wake 5 preprint. Panel a, throughput in transactions per second: Foundry 30,186, Wake 5 29,931, Echidna 2,600. Panel b, distinct assertions violated per million transactions: Foundry 0.32, Wake 5 0.37, Echidna 3.36.
Preprint Figure 2. Medians over 15 twenty-minute runs per tool on the L32 mazes, one physical core each.

Wake 5 executes 29,931 transactions per second, against 30,186 for Foundry and 2,600 for Echidna. The two tools that embed revm run within one percent of each other, so a Python test framework no longer pays for its language in execution speed.

Line chart from the Wake 5 preprint: distinct assertion violations summed over five mazes against campaign time up to eight hours. Medians over five seeds end at 75 for Wake 5, 71 for Foundry and 70 for Echidna, with seed-range bands. A dashed magenta step shows the August agent-written harness reaching all 153 assertions; triangles mark four further campaigns completing all 153 within 80 minutes.
Preprint Figure 3. Fuzzers: medians over five seeds of eight hours each. Agent times include writing the harness and solving.

Throughput does not decide what random fuzzing finds on these mazes. After eight hours the three fuzzers reach 75, 71 and 70 of 153 assertions. A harness written by an agent, with the guards handed to a constraint solver and the routes planned through the parsed maze, reached all 153 in every one of five campaigns, in 33 to 79 minutes. The limit sits in the input generator, not the engine: Wake places half of its 64-bit draws below 2³², which helps on these small constants and loses its advantage when the constants span the full 64-bit range.

The preprint’s static-analysis ablation, where three models reviewed a production Go client with and without an eleven-tool MCP server, found precision up for all three models and recall improved for none: Static Analysis MCP Tools in an AI Audit Benchmark. On real protocols, agent-written tests reached the known recall of hand-written harnesses and still missed the one scenario a known bug needed: Evaluating Agent-Written Fuzz Tests.

Availability

Wake is open source under the ISC licence at github.com/ack3-ai/wake, with documentation at ack3.ai/wake/docs. The preprint is on arXiv under CC BY 4.0: arXiv:2611.09307. Benchmark code and campaign data are at github.com/ack3-ai/daedaluzz-benchmark.

Contributors

About the authors.

Vocabulary

Terms used in this article.