AI Agents Exploit Smart-Contract Bugs, OpenAI Sounds “Code Red,” Apple’s AI Shakeup, and Trump’s Chip Bet
This Week in AI Newsletter: 12/2/2025
Anthropic found its AI agents can exploit real smart-contract bugs, programs on blockchains that move money automatically, generating about $4.6M in simulated exploits on contracts hacked after March 1, 2025. Their paper shows that making money with automated exploits is possible today, so teams should use AI to test and defend contracts before they ship. They created a new benchmark which can evaluate AI agents’ ability to exploit smart contracts called SCONE-bench. More here.
OpenAI CEO Sam Altman has reportedly declared a “code red” push to improve ChatGPT, focusing on personalization, speed, and reliability, even if it means delaying other initiatives. The urgency comes as Google’s latest Gemini release gains momentum and tops some benchmark tests, tightening the race for AI leadership. More here.
Apple is replacing AI chief John Giannandrea, who will remain through spring as an adviser, with Amar Subramanya, a Microsoft executive who previously spent 16 years at Google. The move comes as Apple Intelligence and Siri have struggled with missteps and delays, putting pressure on Apple to catch up in AI. More here.
French AI startup Mistral rolled out Mistral 3, an open-weight lineup spanning 10 models and led by a multimodal, multilingual frontier release. The drop also includes nine smaller models designed to run offline and be easily customized for fast, efficient enterprise deployment. Mistral raised $2.7 billion to date valuing them at $13.7 billion. More here.
Gradium, the French AI voice startup spun out of French AI lab Kyutai, has emerged from stealth with a $70M seed round led by FirstMark Capital and Eurazeo, with investors including Xavier Niel, DST Global Partners, and Eric Schmidt. The company is building ultra-low-latency, multilingual audio language models aimed at making voice experiences faster and more accurate for developers. More here.
Nvidia unveiled Alpamayo-R1, an open reasoning vision-language “action” model aimed at autonomous driving research, plus a set of new developer resources it calls the Cosmos Cookbook. The company says these tools are designed to speed progress toward Level 4 autonomy by improving how vehicles reason about complex, human-like driving decisions. More here.
The Trump administration plans to invest up to $150 million in xLight, a semiconductor startup led by former Intel CEO Pat Gelsinger, that would likely make the government its largest shareholder. xLight aims to improve extreme ultraviolet lithography by building powerful “free electron lasers” to enable smaller, more efficient chipmaking, with a target of producing its first silicon wafers by 2028. More here.
Michael and Susan Dell have pledged $6.25 billion to deposit $250 into investment accounts for 25 million children, aiming to jumpstart their financial futures. This donation is designed to encourage families to participate in the new ‘Trump Accounts’ initiative, a federal program that separately provides a $1,000 government-funded starter deposit for eligible newborns. More here.