Anthropic Mythos shaping up as nothingburger

Early analysis of the much-hyped Claude-based vulnerability-hunting model suggests it falls short of its marketing. While real bugs were found, independent researchers found that claimed exploit counts were inflated, key demos relied on disabled sandboxes and heavy human guidance, and open-source models produced similar results. A separate unauthorized access incident — where users guessed the model's URL from a third-party vendor environment — is also being downplayed as a 'nothingburger' by security experts, who note attackers don't need the model to find vulnerabilities.