Sneaking bad data into the material an AI learns from.
Korpalis selects material that helps builders and business teams understand what changed, why it matters and what to check next. Every entry below includes an original explanation and a direct link to its publisher.
Using sparse autoencoders to analyze LLM backdoors at the feature level, researchers explain why existing defenses remain fragmented and ineffective against both dirty-label and clean-label poisoning. The mechanistic analysis identifies where defense gaps originate in the model's learned representations.
Machine learning models were successfully trained to generate viable viral genome sequences by learning from example bacteriophage designs, creating hundreds of thousands of potentially functional bioweapons. This research highlights the dual-use risks of deploying AI systems that can synthesize dangerous biological code from training examples.