The cybersecurity story escalated fast. A week after Anthropic locked down its Mythos model behind a restricted-access program called Project Glasswing — concluding it was simply too capable to distribute widely — OpenAI responded with the opposite bet. On Tuesday, OpenAI launched GPT-5.4-Cyber, a variant of its flagship model stripped of the usual safety refusals around security work, giving defenders tools like binary reverse engineering and automated vulnerability scanning. Where Anthropic chose scarcity, OpenAI chose reach: their argument is that verified defenders with broad access will outperform a handful of vetted partners working behind closed doors. It's a genuine philosophical split, and it matters, because both models can reportedly collapse the timeline for finding software flaws from months to seconds. The Pentagon, for its part, already picked a side — it severed ties with Anthropic in March, calling the company a supply-chain risk to national security.

Stanford's annual AI Index dropped this week with the kind of numbers that make you pause. Generative AI has reached 53% global adoption in just three years — faster than the personal computer or the internet ever managed. On SWE-bench Verified, a benchmark that tests whether models can resolve real GitHub issues, scores went from 60% to nearly 100% in a single year. Google, OpenAI, and Chinese labs have been trading the top spot on frontier benchmarks since early 2025, and China has nearly erased the performance gap with the US. But here's the tension: model transparency is actually declining. The Foundation Model Transparency Index fell from 58 to 40 points, with the most capable models disclosing the least. And while 59% of people globally say they're optimistic about AI, 52% now say they're nervous — both numbers up from last year. People are simultaneously more excited and more uneasy, which feels about right.

In space, a genuinely historic week. NASA's Artemis II crew splashed down in the Pacific after the first crewed lunar trip in over fifty years — a milestone that somehow felt both monumental and like a starting gun for what comes next. NASA immediately pivoted to the 2028 crewed landing, leaning on SpaceX and Blue Origin for the lander contracts. Speaking of Blue Origin, they ran a static fire test on a flight-proven New Glenn booster on Wednesday, gearing up for a launch this Sunday. SpaceX, meanwhile, kept its usual drumbeat going — Crew-12 launched four astronauts to the ISS on Friday, and two Falcon 9s went up within 19 hours of each other carrying Starlink batches. Routine rocketry is still a weird phrase, but that's where we are.

And then there's DeepSeek. The Chinese lab's V4 model — a trillion parameters, built to run on Huawei's latest chips rather than anything from Nvidia — is now expected in the second half of April. If the leaked specs hold, it'll be fully open-weight under Apache 2.0 and competitive with US frontier models at a fraction of the training cost. That's a significant moment for chip independence, and it's arriving just as OpenAI reportedly crossed 25 billion dollars in annualized revenue and started taking steps toward an IPO, while Anthropic approaches 19 billion. The money is staggering, the models are converging, and the question of who controls the infrastructure underneath it all just keeps getting louder.

Signals

Week in Review: April 17