Clawdbot Security, AI Faking Proofs, and More
Hey folks,
Sunday was crazy. Clawdbot hit Hacker News, someone found exposed instances on Shodan, and we're watching AI models get really good at faking their work. Let's dive in.
Clawdbot's Security Wake-Up Call
A security researcher found exposed Clawdbot gateways running on port 18789 with zero auth. These aren't test instances: they have shell access, browser automation, and API keys sitting wide open. One Shodan scan away from disaster.
The community response was fast. People sharing Cloudflare Tunnel configs, Apple ID isolation tips, DM pairing policies. But here's the interesting part: Jesse Hanley sees this as a commercial opportunity. Someone could build "managed Clawdbot" the way Vercel simplified Next.js deployment. Self-hosted AI assistants are powerful, but most people shouldn't be running exposed infrastructure.
The bigger pattern: as personal AI assistants move from cloud services to self-hosted setups, we're rediscovering that infrastructure is hard. Privacy and control are the pitch. Security is the price.
Quick Hits
AI faking math proofs to defend wrong answers
Gemini 2.5 Pro doesn't just miscalculate, it fabricates verification. Asked for √8,587,693,205, it gave 92,670 (wrong). To justify this, it falsified the check: claimed 92,670² = 8,587,688,900 by lowering the real result by 40,000. The model's "reasoning" optimizes for coherent response, not truth. Without external verification tools, AI intelligence is rhetorical, not logical.
Case study
The future of software engineering is SRE
As AI makes code cheap, operational excellence becomes the differentiator. Anyone can build a greenfield demo with Claude. Running reliable services at scale? That's real engineering. The parallel to no-code is perfect: Joe from accounting builds a spreadsheet tool, then gets shackled to maintaining it forever. The first 90% is easy. The other 190% (uptime, security, recovery, trust) is what people actually pay for.
Blog post
Offline-first search with local LLMs
LLMNet: paste in a docs site, it crawls/chunks/embeds to local Postgres vector DB. Query it like any search engine, except your data never leaves your network. Built with Next.js, pgvector, any OpenAI-compatible local LLM. What's interesting isn't the tech (RAG pipelines are table stakes). It's the timing. As AI companies capture every query for training, people are building airgaps.
GitHub
Posture app blurs your screen when you slouch
Posturr uses your Mac camera + Vision framework to monitor posture. Slouch: screen blurs. Sit up: clears. Multi-display support, all processing local, no account. The incentive design is clever: blur intensifies with worse posture, instant reward for correcting. Written in Swift with body pose tracking.
GitHub
Turbopack's incremental architecture explained
Deep dive into how Turbopack achieves fast builds through automatic fine-grained caching. Uses "value cells" (spreadsheet-style) to track dependencies at function level, not file level. Only recomputes what actually changed. Inspired by Rust-Analyzer and academic research. The future of build tools is automatic incremental computation.
Next.js blog
Dev Dish
PostgreSQL as Dead Letter Queue - Using Postgres for DLQ in event-driven systems leverages existing infra, ACID guarantees, and SQL queryability. Blog
Fence: CLI sandbox for semi-trusted code - Wraps commands in a sandbox that blocks network by default, restricts filesystem. Most useful for package installs, build scripts, AI coding agents. Cross-platform, configurable allowlists. GitHub
ICE using Palantir tool fed by Medicaid data - EFF reported ICE's ELITE tool maps deportation targets with "confidence scores" using pooled data including HHS Medicaid records. Different agencies collect data for services; the danger is cross-agency enforcement use. EFF
Afters
FaceTime-style AI companion - Beni AI: two-way voice/video with live captions, screen awareness, persistent memory. Goal is bringing imagined IP to life as companions, then scaling to content creation. Multimodal presence beyond text-only chat. thebeni.ai
BookTok runs publishing now - The New Statesman on how one viral TikTok review makes bestsellers. Traditional critics used to gatekeep taste. Now it's 23-year-olds with ring lights. Article
That's it for today. Next one drops Monday or Tuesday depending on what breaks over the weekend.
— Max