Anthropic frontier red-team lead Logan Graham discusses AI cybersecurityNetwork security protects communications infrastructure and connected systems from unauthorized access, disruption and misuse. with MTS host Theo Jaffee, including the uncertainty around future attacks and defenses.
Graham describes a controlled proof of concept involving small models and self-replicationA self-replicating agent is a software agent capable of creating or initiating additional instances of itself.. He acknowledges criticism of the experimental conditions and says current small-model worms are slow and not presently useful, while arguing the work can reveal future risks.
He discusses defensive access, the practical bottleneck of fixing vulnerabilitiesPatch management is the organised process of identifying, evaluating, deploying and verifying software updates that fix defects or security weaknesses., Anthropic’s Project Glasswing and his views on biosecurity. Predictions and assessments are attributed to Graham.
Watch on YouTube
