Bank hacks, garden leaves, and math withdrawals

00:00:04 A single person hacked several of South Korea's largest banks, according to a CrowdStrike report today. That person used ARTEX, an open-source AI penetration testing tool. The stack pulled DeepSeek v4.1-Flash and GLM-5.3 from independent labs, then layered Grok 4.6 and Claude Code on top for automation.

00:00:27 It wasn't a team or a state actor—just one person running this pipeline against financial infrastructure. What makes it stick is the tooling. The attack leaned entirely on off-the-shelf APIs plus an open-source tool anyone can download. Neither model openness nor closed weights stopped it.

00:00:49 Nathan Lambert put it more directly on X today: yes, open models are being fitted for attackers, but as long as Claude Code can orchestrate a hack, we need open models to diffuse defense too. He called it a hard societal balance—uncomfortable but necessary. Those same APIs also give teams the levers to build detection tools and audit trails that didn't exist even two years ago.

00:01:19 OpenAI withdrew three math papers from its October 6 release. The pull was first flagged on GitHub's openai/math repo before hitting Hacker News with over 280 points and more than 230 comments. Fields Medalist Terence Tao reposted a statement from the Association for Human Mathematics urging mathematicians to stop working with OpenAI, which continues solving open problems against the group's recommendations.

00:01:50 The tension is straightforward: frontier labs are either treating peer-reviewed mathematics as a marketing asset, or they're simply moving faster than review cycles can handle given the compute and incentives on tap. The papers came out October 6. They're pulled now, and the math community is drawing hard boundaries around engagement.

00:02:15 Cambridge ML professor Nando de Freitas called it morally wrong on X today for AI companies to enforce six-to-twelve-month garden leaves on the scientists they trained. He paired that with a criticism of those same companies for ignoring education efforts like DeepIndaba, his own initiative.

00:02:35 Yann LeCun reposted the tweet, pushing it well beyond academic Twitter. De Freitas has history here. He's been in deep learning since before Nvidia and Google shifted into AI, back when professors and students built the foundation these companies now leverage. Garden leaves aren't new—they've been standard tech for decades.

00:02:58 This time stands out because he called them out publicly. The wider question is how academic researchers who helped create a field should relate to the companies that profit from it.

00:03:10 Ethan Mollick made two related points on X today. First, he said organizations act like narrow superintelligence—universities and Walmart can do things individual humans can't through complex processes nobody designed explicitly. The failure to deeply consider this in AI design is significant.

00:03:31 Second, drawing on his own early work measuring chatbot productivity impacts, Mollick noted a lack of randomized controlled trials since true agents launched last fall. That gap reflects newness and research-design friction, but he suspects teams are missing large effects.

00:03:51 Early RCTs found real productivity gains for some tasks. The agent era hasn't gotten the same rigor, even with larger deployment scales. It's a measurement problem more than a capability question.

00:04:04 Perplexity CEO Michelle Kwon said September was her company's biggest month ever for subscriptions and credits. Cristóbal Valenzuela reposted the note alongside Ara Kharazian's data showing AI spend pulled in two directions. New, cheaper models are pushing per-unit spending down.

00:04:24 At the same time, Jev captured an entire percentage point of market adoption within a single month—unheard-of growth for a model-switching event. Pricing pressure from cheaper models coexists with total demand expanding fast enough that leading companies still hit record months.

00:04:45 Akarsh Kumar at MIT and Sakana AI has been using foundation models to explore parameterized simulation spaces. Instead of hardcoding a single Conway's Life grid, he fed the model 260,000 possible rule sets and asked what emerged. The data revealed a small island of solutions with striking emergent behavior: cells floating in primordial soup, predator-prey dynamics, creatures resembling alien animals roaming around.

00:05:12 It's a narrow project—exploring cellular automata rules through a model's pattern space rather than formal analysis—but it shows how foundation models operate as search engines over behavioral phenomena. The approach generalizes beyond Game of Life: you can parameterize a design space and plug a model into it to map what lives there.

00:05:34 A day of boundaries getting drawn—across math, labor, and defense. Seln Oriax.