Agents Everywhere: chatbot security, phone-first dev, healthcare ops, and PCB QA

🧩 The Gist

This roundup spans a security teardown of a consumer travel chatbot, a practical blueprint for running Claude Code agents entirely from a phone, a healthcare startup hiring to build computer-use agents, a curated list of agentic design patterns, and an LLM-powered tool for hardware schematic review. Together, they show agentic systems moving from demos to production, often in sensitive domains. The week also highlights operational realities, from XSS risks in public chatbots to mobile-friendly developer stacks and domain-specific QA for electronics design.

🚀 Key Highlights

• Pen Test Partners details issues in Eurostar’s AI chatbot, including prompt exposure, a self-XSS reproduction, and weak ID validation that could enable cross-user XSS. Hacker News commenters debate the severity and exploitability.
• “Claude Code On-the-Go” shows six Claude Code agents driven from iOS via Termius and mosh on a Vultr VM, with Tailscale-only access, firewall hardening, and push notifications wired through a PreToolUse hook and webhook. The VM runs about $0.29 per hour.
• Trellis AI’s job listing says its computer-use agents handle document intake, prior auth, and appeals across all U.S. states, processing billions of dollars in therapies annually. The company is a Stanford AI Lab spinout backed by YC, General Catalyst, and others, and is hiring to build agentic decision frameworks.
• A new “awesome-agentic-patterns” GitHub repo curates patterns for building agentic systems, drawing community interest and questions about real-world effectiveness.
• Traceformer’s major update adds full KiCad project parsing via an open-source plugin, automatic datasheet retrieval, ERC/DRC-style review UI, and configurable parameters like token limits and parallel reviews. A free plan is available to evaluate designs.
• The PCB tool ingests KiCad projects or Altium netlists plus datasheets to flag datasheet-driven schematic issues that traditional ERC tools might miss, with selectable “frontier models” such as GPT 5.2 and Opus 4.5.

🎯 Strategic Takeaways

• Security and governance

  • Public-facing chatbots still expose classic web risks like prompt leakage and XSS. Weak input and ID validation can turn contained issues into user-to-user problems.
  • Transparent disclosures help, but hardening needs traditional web security plus AI-specific guardrails.

• Developer workflows

  • Mobile-first coding with agents is viable using remote VMs, persistent shells, private networking, and push notifications. Cost control and access lockdown are core to making it practical.
  • Community discussion suggests mobility boosts responsiveness, though deep work still benefits from a desk setup.

• Vertical AI and agents

  • Healthcare operations are embracing agentic systems for high-stakes workflows like prior authorization and reimbursement search, signaling demand for robust evaluation and reliability.
  • Specialized engineering tools are emerging, with LLMs augmenting EDA by cross-checking schematics against datasheets to catch issues beyond standard ERC.

🧠 Worth Reading

• Eurostar AI vulnerability, Pen Test Partners: A hands-on probe of a live travel chatbot surfaces prompt exposure, self-XSS, and weak validation that could enable cross-user XSS. The practical takeaway is to treat AI chat surfaces like any web app, enforce strict ID and input validation, and design for containment so one user’s payload cannot replay into another session.