Field Notes

Security leak in ChatGPT, new agentic benchmark GDPx, three actionable agent tips

Today is about security, better model choice by domain, and practical agent workflows.

Key AI news

ChatGPT: major data leak in HTML previews fixed

Researchers at Parameter Labs took over accounts through a chain of sandbox bypass, iFrame injection, and JavaScript execution in ChatGPT’s HTML previews. OpenAI used MathLive. Its virtual keyboard sent HTML to the parent page via postMessage. That allowed insertion of an unsandboxed iFrame. The Content Security Policy (CSP, a browser rule for allowed resources) allowed YouTube. Its oEmbed endpoint reflected script code. Private data could be taken or actions triggered in the name of logged-in users. OpenAI fixed the hole and paid a 5,000 dollar bug bounty. For you: open shared ChatGPT links only in a separate browser profile or while logged out.1

GDPx: new benchmark for agentic skills in value-creating domains

SemiAnalysis introduces GDPx. A benchmark across chip design, cyber, bio, finance, law, quant, RSI, hardware, materials science, alignment, and software. According to the provider the best model reaches 12 percent overall. Strong domain differences and a clear impact from safety classifiers. These are filters that classify risky content. Opus 5.5, GPT-6 Astra, and Gemini 4 Carbon lead by area. Muse Wonder Full is ahead in cyber. Consequence: you pick models by your domain, not by a single overall score. Note: results and framing come from the provider and partners.2

Agents in the wild: SETI steers telescopes with natural language

Breakthrough Listen and the SETI Institute use agentic systems at the Allen Telescope Array. Through Holoscan, terabits per second go straight into GPU memory. CyberEther processes the signals. Nexus serves as web-based agent orchestrator. The facility hits 2.4 Tbit/s on RTX PRO 6000 Blackwell and detected the Crab pulsar in real time. For you: this is a blueprint pattern for your own sensor pipelines with live correction when latency and data rates matter.3

Highlights for your workday

Arca in finance shows how agents support longer processes without automating money movement. Agents analyze and propose. Humans decide deterministically and check prompt injection. A persistent customer memory holds tasks, ownership, and deadlines. Take a 10-day process. Give the agent analysis and reporting. Keep payments with humans.4

HackerTube curates YouTube from 40,000 HN comments and 6,454 videos instead of clickbait. Local session profiles, your own recommendations without an account, and data export are built in. It was built in a few hours with Opus 5.5. You can build a sober video start page for team learning or feed your agent with the export.5

Test Copilot and others realistically without real company data. Microsoft describes synthetic tenants. The environments draw on O*NET, job ads, and SEC data. They include mail, meetings, documents, and timelines. You find weaknesses that generic benchmarks miss. Start small and measure accuracy, not vibes.6

Tools and updates

  • Apriel Voice 27B Preview: open speech model for understanding, dialog, and tool calls, also text and image. Use: voice assistant prototypes with tool use. Requirement: compute resources and ML know-how, it is a preview.7
  • NVIDIA: deploy models at the edge with a few commands. Useful for robotics, vehicle cabins, and manufacturing with tight latency and power limits. Requirement: NVIDIA stack on edge hardware.8
  • Ropes & Gray: AI-assisted transaction due diligence with OpenAI. Lawyers review grouped documents, summaries, and reports before client counsel. Usable as a template for legal teams, access to OpenAI required.9

Try it in five minutes

Mini-eval without real data. Check if your chat assistant summarizes and cites meetings cleanly.

1) Create sample data in a document: - Three emails: “Sales reports 12% Q4 revenue increase”, “IT plans CRM update by 11/15”, “Legal reviews new T&C clause by 11/08”. - Agenda: “Q4 outlook, CRM timeline, T&C risks”. - 10-line transcript with two clear decisions and two open points. For example: “We move the launch to 11/20.” and “Marketing delivers landing page draft by 11/06.”

2) Prompt in your chat tool: “Summarize the meeting in five sentences. List three decisions and two open points. For each point, quote the exact sentence from the data.”

3) Expected result: a tight summary with five entries. Test criterion: all decisions and open points are correct, and the quotes appear verbatim in the material. If quotes are missing or facts drift, mark the errors and adjust your prompt.

This idea follows Microsoft’s approach of using synthetic environments for realistic evaluation.6

Sources

  1. How We Took Over ChatGPT Accounts (simonwillison.net)
  2. GDPx: Measuring Frontier Model Capabilities Across the Most Valuable Domains (newsletter.semianalysis.com)
  3. Things That Go Beep in the Night: Agentic AI Joins the Search for Extraterrestrial Signals (blogs.nvidia.com)
  4. Agents Prompt People: Using Agents for Long-Horizon Financial Work (latent.space)
  5. HackerTube (karpathy.bearblog.dev)
  6. From synthetic enterprise data evaluations to real Microsoft Copilot AI confidence (microsoft.com)
  7. Paper: Apriel Voice 27B Preview: An Open Model for Speech, Reasoning, and Tool Use (huggingface.co)
  8. Run AI Models for Edge Application Development with Just a Few Commands (developer.nvidia.com)
  9. How Ropes & Gray is developing AI for transactional diligence (openai.com)