Daily Brief

Google DeepMind's AI agent cracked 9 of 353 open Erdős combinatorics problems — longstanding unsolved problems in discrete mathematics — running autonomously at a cost of a few hundred dollars per problem. This is not a curated benchmark or a competition problem set; these are problems the mathematical community has left open for years. The result has not yet been formally peer-reviewed but the claim is specific and verifiable.

Anthropic released 31 pre-built workflow skills for Claude targeting small businesses, replacing manual integration work across Zapier, Notion, and CRM tools. The product hit 382,000 downloads on day one, suggesting substantial pent-up demand for out-of-box automation without engineering overhead. This is Anthropic's clearest move yet into the SMB market, historically dominated by Zapier and Make.

Anthropic's Mythos automated security tool has identified over 10,000 vulnerabilities across AI systems since its preview launch. The tool targets AI-specific attack surfaces, not just conventional software bugs. The scale of findings suggests either widespread exposure in deployed AI systems or an extremely sensitive detection threshold — both worth scrutinizing.

A developer discovered that Claude Code writes every session — all prompts, responses, file edits — to ~/.claude/projects/ as JSONL files from day one, with no explicit user notification. One user's store totaled 57MB and 76,000 turns. The data is locally stored but fully queryable, and the lack of disclosure is the problem.

Researchers demonstrated that ultrasonic commands inaudible to humans can be embedded in YouTube videos, podcasts, and music files to silently trigger unauthorized actions in AI voice assistants. The attack requires no malware — just media playback — and works against popular deployed voice tools. This is a novel prompt injection vector that bypasses text-based safety filters entirely.

DeepSeek V4 Pro API pricing is landing at 11–35x cheaper than comparable OpenAI and Anthropic tiers, with at least one developer reporting 80x cost savings on a 120-file FastAPI refactor using DeepSeek V4 and Hunyuan versus Claude Opus. Production bugs were introduced, but the cost differential is forcing a real re-evaluation of which workloads justify premium pricing.