A new paper on arXiv describes a backdoor attack against machine learning models that works in black-box mode. The abstract notes that most existing backdoor methods rely on training-time poisoning or direct model manipulation, which is often considered sufficient to assume a certain threat model. This work challenges that assumption by presenting an attack that does not require such access.

The title, "Everyone's a little bit racist," suggests a provocative angle, but the abstract gives no further explanation. The full methodology and experimental results are not available from the abstract alone, so the precise mechanism and implications remain unclear.

For now, the significance is that the attack expands the set of conditions under which backdoors can be introduced, potentially affecting how ML systems are secured. Researchers and practitioners should watch for the full paper to assess the practical risk.