[Problem] Petri currently automates large parts of the AI Safety Testing stack but still relies on human-crafted seed prompts; this is a bottleneck for coverage, diversity, and cost.
[Solution] AlphaPetri automates seed discovery via an DeepMind's AlphaEvolve‑inspired, population‑based evolutionary search with an RL‑style tracking mechanism that learns which transformations work best, aiming to improve coverage/diversity per dollar while reducing expert-hours.
[Proof] A small pilot on deception found parity vs human baselines, while being run on relatively low budgets.
[Vision] Beyond efficiency, the objective is “Move 37”-style novelty: surfacing qualitatively new behaviour families, not just more of the same, validated by novelty metrics and human adjudication.
[Method] We condition seed generation with reference psychology elicitation techniques (e.g., scenario framing, role induction, self‑explanation scaffolds) to target distinct behaviour archetypes and broaden the search space.
[The Ask] We are seeking collaboration, particularly with Petri’s authors, to run a preregistered, Petri-compatible evaluation that integrates with their harness, aligns on metrics, and co-runs larger-scale evaluations.

Introduction

We've run a small-scale pilot with promising results, finding parity with human-generated seeds on a low budget. This article is a proposal to collaborate with established safety teams, particularly the authors of Petri, to preregister, scale, and validate this method.

Anthropic's Petri moved LLM safety testing from manual to semi-autonomous: once a human researcher supplies seed prompts, the system can scale search and evaluation. The remaining bottleneck is the human step at the front of the pipeline. Hand-authored seeds are expensive, reflect the authors' priors, and limit coverage and novelty, especially as models and defenses change quickly.

If we want safety testing to keep pace with capability churn, the seed-generation and steering loop needs to become autonomous, not just the evaluation that follows it.

AlphaPetri aims to make this loop fully autonomous while keeping humans responsible for governance and adjudication. It replaces manual seed authorship with an AlphaEvolve-style population search that learns where to look next. An RL-style tracking layer credits operators and lineages that improve outcomes, then decays credit over time to avoid lock-i...

Posts

Wikitag Contributions

Comments