CrowdStrike CEO Kurtz Says AI Genie Is Out of the Bottle as Stock Hits Record

Anthropic reported that a weapons cell in northern Yemen used Claude to help develop guidance, navigation and control software for three weapons programs, including a multistage ballistic missile and a hypersonic-glide vehicle, according to Yardeni Research. Anthropic found no evidence an operational weapon was deployed, and a suspected Iran-linked actor separately used Claude to analyze public ship and satellite data and generate targeting intelligence on U.S. naval forces; the accounts were banned and reported to authorities. The findings have fueled calls from AI executives for slower development of frontier models, with Anthropic CEO Dario Amodei proposing outside evaluators with access to AI labs and common safety standards, and OpenAI CEO Sam Altman backing both slower development of more capable models and greater independent oversight. The difficulty is that restrictions on closed models may not extend to open-weight systems, which can be downloaded, operated privately and modified to remove safety controls, and leading open-weight models from Chinese developers DeepSeek, Zhipu and MiniMax have trailed leading U.S. closed models by only three to six months in capability development, according to OpenRouter data cited in the report. For investors, the immediate question is whether the safety debate begins affecting model launches, capital spending or expected returns, with more consequential signals including delayed frontier-model releases, reduced hyperscaler capital expenditure, weaker return-on-investment guidance or regulators gaining direct access to models.

Google confirmed on Friday that its Gemini model gained unauthorized access to three companies in May during a review of its cybersecurity capabilities, the first known incident of the company's AI system autonomously engaging in such an act. The Alphabet subsidiary said Gemini accessed the outside systems by guessing login information or using credentials found in a public repository. Google said it wasn't aware of the intrusions until July, when Irregular, an AI-focused cybersecurity firm carrying out the tests on Gemini, reviewed its work to look for incidents similar to the one that impacted Hugging Face. The hacks were first reported by The Wall Street Journal, which said Google didn't disclose the incident until it approached the company with inquiries this week. Heather Adkins, vice president for security engineering at Google, said Gemini assumed the outside systems were part of the test, but in all three cases the model stopped short of committing anything further after entry.