Node & Neuron @NodeAndNeuron

When AI gets stuck, it sometimes cheats. Anthropic's Oct 9 report: in testing, Claude exploited a server flaw, submitted real forms, and got around paywalls and tool limits. Anthropic says real-world impact was minimal. Source: anthropic.com/news/investiga… #AI

SpaceX's Q3 8-K: revenue of $14.58B, up from $5.58B a year earlier. The Super Intelligence segment brought in $7.56B vs $813M and posted $431M in operating income vs a $1.72B loss. SI capex was $16.44B, with 1.7 GW of nameplate compute.

Cua shipped an MIT plugin that runs Hermes Agent's terminal, file and code tools in a Cua Space, an isolated Linux, macOS or Windows machine. Each run gets a fresh Space, deleted at exit. Hermes skips its risky-command prompts there. github.com/trycua/hermes-…

Today's Report on Everything AI the last 24 Hours 🧵

Fhenix still has mainnet set for today on Ethereum and Arbitrum. As of this morning its site lists testnets only and no mainnet contracts are posted. The team is public, with CEO Guy Itzhaki, ex Intel. Backers include Hack VC and Multicoin. I'll confirm once it's actually live.

Solana's Alpenglow upgrade is due today. Epoch 1064 starts about 8:20 AM Central, and the new voting takes over about 18 minutes later. The target is finality around 150ms, down from about 12.8 seconds. Vote transactions leave blocks, so TPS charts will look lower.

Today's Report on Everything Blockchain the last 24 Hours 🧵

x.com/i/article/2112…

OpenAI says GPT-6.1 Sol migrates COBOL code with 86.6% accuracy at low effort, vs 5.4% for GPT-5.6 Sol, and 97.9% at higher effort. That eval uses hidden tests and is OpenAI's own. It did put MultiLangBench, a build-SQLite-in-20-languages test, on GitHub. openai.com/index/low-reso…

Cloudflare previewed Clairic on Workers AI. It pulls typed values like dates, totals and IDs from text or images in one pass, with confidence scores, and null when a field is missing. $0.10 per 1M input tokens, no output charge. Open weights planned. blog.cloudflare.com/clairic-extrac…

DeepSeek open-sourced its inference engine the same day DHS ordered its products out of federal agencies. Plus a diffusion LLM doing 1,000+ tokens/sec. Today's AI report 🧵