A summary of the HaloGuard 2.0 paper: a small open streaming guard that comes out ahead of the current open streaming guards, the shortcut it had to unlearn, what the paper reports, and what it does not claim.
Read articleA summary of the HaloGuard 1.0 paper: how a written constitution of 46 policies, paired counterfactual prompts and balanced multilingual data were used to train a small input guard, what the paper reports, and what it does not claim.
Read articleA distributed adversarial research network of 392 independent red-teamers produced 200 structured alignment audit scenarios, published as an open dataset for Petri-based AI safety testing.
Read articleA proposal and pilot study for automating LLM safety testing using evolutionary search to generate seed prompts, addressing bottlenecks in human-crafted seed prompts for Petri's testing harness.
Read article