Augustus Forge
Why
Scanners and one-shot prompts find what everyone else finds, and the duplicates pile up. A senior researcher works differently: build a model of the system first, then hunt the gap between the model and the code. The toolkit makes an agent work that way every time, whether or not it remembers to.
What came out
A self-driving hunt loop: pick one hypothesis, drive it deep, prove it or kill it, repeat. The discipline lives in runtime hooks, not in the prompt: a ledger that exists before the first action, a loop that won’t let the agent quit early, a prompt-injection guard over everything external.
Nothing is reported until a fresh-context agent re-derives the finding and a separate check confirms scope and payout.
What broke
- Not every lead holds up. Cold verification kills roughly a quarter of the candidates before they become a submission.
- Parts of the internals are still in Russian. Much of it is format strings the runtime matches on, so it gets translated slowly, without breaking it.
- Four self-tests need a private benchmark corpus that isn’t in the public release.