Banks warn on Claude Mythos & AI agents write full papers - AI News (Apr 11, 2026)

[0:36] Banks warn on Claude Mythos

Cybersecurity and policy first. Reports say U.S. Treasury Secretary Scott Bessent convened leaders from major banks to discuss risks tied to Anthropic’s newest model, Claude Mythos, with the Federal Reserve’s Jerome Powell also said to be present. The worry is simple: if AI meaningfully boosts vulnerability discovery and exploitation, it doesn’t just raise the baseline for hackers—it raises the baseline for systemic incidents across payments, identity systems, and core banking infrastructure.

Consumer AI is also moving into more sensitive territory. Perplexity expanded its Personal Finance experience with Plaid connections, letting users link accounts and ask natural-language questions about spending, liabilities, and net worth. The appeal is clear—one dashboard, one conversational interface. The hard part is trust: giving an AI assistant a complete financial picture raises the stakes on security, data handling, and the possibility of subtle mistakes becoming real-world consequences.

On the tools side, a notable library update: Sentence Transformers added multimodal embedding and reranking support, aiming to make cross-modal search—text-to-image, text-to-video, mixed retrieval—feel like a straightforward extension of existing APIs. This matters because a lot of AI product work is quietly becoming “find the right thing in messy data,” not “generate new text,” and multimodal retrieval is increasingly central to that.

In generative media research, two different papers point to a similar direction: making generation more controllable and scalable. One proposes process-driven image generation that iterates through planning, drafting, critique, and refinement—closer to how humans draw—so the model can correct itself over multiple steps. Another, Sol-RL, aims to make reinforcement learning style alignment for diffusion models cheaper by using low-precision rollouts for selection and higher precision where training stability matters. The shared theme is less mystique, more discipline: explicit iteration, tighter feedback loops, and lower training cost.

Google also pushed interaction over static answers: the Gemini app can now generate interactive simulations, dynamic charts, and 3D models inside chat, so users can manipulate parameters and see outcomes. If it works reliably, it’s a meaningful upgrade for learning and exploration—because understanding often comes from poking at a system, not just reading about it.

One more perspective piece worth your time: Quanta Magazine argues that many viral ‘AI horror’ anecdotes become scarier by omitting the human instructions that shaped the behavior. The point isn’t that risks don’t exist—it’s that the most urgent problems may be more mundane and more immediate: misinformation, over-trust, and people delegating judgment in settings where an LLM can sound confident while being wrong.

Finally, a reality check on long-horizon agents. KellyBench evaluates models in a simulated sports betting market over an entire season, forcing sequential decisions, risk management, and adaptation. The headline result: every tested model lost money on average, and many went bankrupt. It’s a crisp reminder that sustained strategy under uncertainty—staying consistent, updating beliefs, sizing risk—remains a major weak spot for today’s frontier systems.