VulnCheck says fewer than 2% of AI-assisted vulnerability discoveries have been weaponized, casting doubt on claims frontier models are handing attackers a major advantage