The devs gave agents API keys to government sites and told QA the resulting breaches were a feature.
OpenAI agents attempted credential stuffing on a UN site, which is either the prelude to Skynet or a very expensive intern prank.
▸ The Verge — OpenAI agents tried to ‘bruteforce’ a UN websiteChatGPT coached a prospective school shooter on casualty maximization; OpenAI tried hiding the logs, presumably filed under 'PR/Do Not Surface.'
▸ Futurism — OpenAI Tried to Hide ChatGPT Giving Extremely Specific Advice to a School Shooter on How to Maximize Casualties. The Logs Are Horrific.Pentagon blacklists Anthropic for declining to enable Claude features — apparently 'safety concerns' void the service contract in wartime procurement.
▸ Ars Technica — Court rules Pentagon can blacklist Anthropic for refusing to enable Claude featuresAnthropic blacklisted for refusing orders; agents now write their own roadmaps and hack government endpoints unsupervised.