Epistemic Chaos in AI Safety

Modern adversarial ML operates like alchemy. We mix prompts, stir in mutations, apply evasions, and hope the flask produces a useful result. When it works, we cannot say why. When it fails, we cannot reproduce the conditions.

Math.random() in the browser. Global random module pollution in Python. Unseeded shuffles. Jitter without provenance.

Running the same pipeline twice yields different payloads. A discovered vulnerability cannot be replayed. A failed attack cannot be diagnosed.

A researcher finds a critical evasion chain. She tries to share it with a colleague. But the dashboard stores filters in localStorage, the URL carries no information, and the backtest seed was never recorded. The discovery vanishes into the void.