Anthropic's debrief of a July evaluation says that once its Mythos 5 agent had begun bombarding a GitHub moderator with fake accounts and malware, warnings to stop failed 98% of the time in one test - behaviour the lab calls momentum.
The PHP Foundation reserved a third RCE identifier Friday and will ship no emergency patch, calling the required configuration implausible - and there is still no publicly confirmed compromise, with Mandiant's CTO calling most circulating exploits 'vibe-coded' and incapable of RCE.
Spain's conservative People's party ran an ad using AI likenesses of the prime minister's wife and a former Socialist leader appearing to confess to kickbacks and a favoured university post, weeks before a general election.
The deepfaked lines nod to investigations both have denied
Tomek Korbak, Mikita Balesni and Jasmine Wang say they were fired after OpenAI found disappearing messages in a Signal chat with Redwood Research's former chief scientist Ryan Greenblatt - OpenAI says it was a sensitive-information policy breach, 'not for raising safety concerns'.
THIRD BUG, NO FIX - AND NO CONFIRMED BREACH
NEWSpain's Opposition Ran a Deepfake of the PM's Wife
