Cursor vendor-key breach, Vercel agent evals, Robinhood engineer guilty plea, Hugging Face and OpenRouter, Island screen-time app
Hello, dear TEA-mates! Here is what you need to know today.
1. 🔐 An Update on Cursor’s Recent Security Incident
The Cursor Team wrote on 13 November 2026 that an attacker obtained an access key Cursor had issued to Fireworks, that Fireworks exposed the key, and that the attacker used it to copy data from a read-only storage bucket Cursor used to share data with that vendor. The bucket held limited coding agent traces, mostly from 2025, from sessions not in Privacy Mode, which Cursor was permitted to use for model training. A small number of those traces contained credentials that had not been redacted, and less than 0.01% of customers had API credentials exposed. Starting with those credentials, the attacker reached select SaaS tools used by Cursor and accessed archived non-production source code repositories, limited customer support data and email conversations, product usage metrics and contact details, summaries of some coding agent conversations, and some credentials users had sent to Cursor support. On 29 October, using a stolen credential, the attacker placed a malicious, unauthorized copy of Anthropic’s Claude Code extension on the cache that serves the Cursor VS Code extension marketplace; Anthropic did not publish it, and Open VSX never had the malware. At most 5,000 users who requested specific cached versions on 29 October potentially received the malicious build, which was available for about 11 hours, and Cursor asked anyone who installed Anthropic.claude-code 2.1.312 or 2.1.313 from that marketplace during the window to follow its security notice. Cursor said the last attacker activity it had seen was on 1 November and was blocked, that it believes the incident is contained, and that for users not in Privacy Mode, coding agent traces were exposed for fewer than 0.05% of users, each of whom was notified. (Read More)
🫖 TEA For Thought: “In this situation, what do you do? Your security guard is safe, but depending on all the other vendors that you give access to, they can still get into your system. A security issue is a global thing, and I think everybody is trying to figure it out.”
2. 🧪 Agents Are Users, So Your Software Needs Evals
George Fahmy and Yury Selivanov of Vercel wrote on 16 November 2026 that agents are now users of software, doing work on people’s behalf, and that whether you are building the agent or the software it uses, such as an API, CLI, MCP server, or SDK, evals are how you measure whether that work gets done well. They argue that an agent can place a “successful” order for a hot latte when the user asked for an iced one, so request logs can look fine while the outcome is wrong, and that because model output varies between runs, one good manual test cannot tell you whether a fix worked. An eval, as they define it, is a test you run many times and score. They group agentic tests by how many model calls they typically need: deterministic tests (none), one-turn evals (one), and end-to-end evals (two or more across a whole task), and say that for “Success” you usually need the top of that pyramid, because each step can look reasonable and the final result can still miss the user’s requirements. Popular coding benchmarks such as SWE-bench and Terminal-Bench are examples of end-to-end evals. The authors walk through building one with Harbor, an open-source framework from the creators of Terminal-Bench, and recommend metrics including success rate, pass@k, pass^k, and, on successful trials only, steps, time, and cost. (Read More)
🫖 TEA For Thought: “Your agents are your users. Running an API successfully doesn’t mean jobs got done and jobs got done well. Have your evals.”
3. ⚖️ Former Robinhood Employee Pleads Guilty to Fraud
The original U.S. Attorney’s Office for the Southern District of New York press release was blocked by a bot check, so this summary is based on independent coverage of that announcement by CoinDesk and The Block. On 16 November 2026, Huaisong Xiang, a former software engineer at Robinhood Markets, pleaded guilty in Manhattan to wire fraud, before U.S. District Judge Colleen McMahon, after using confidential information about upcoming Robinhood Crypto listings to trade perpetual futures on Hyperliquid. According to the office, as reported by both outlets, Xiang made those trades between 2025 and 2026, ahead of Robinhood’s public listing announcements, and agreed to forfeit $125,523 in proceeds. U.S. Attorney Jamie McDonald said, “No matter the trading platform, corporate insiders will be held accountable when they trade based on misappropriated information,” and that Xiang had “repeatedly traded in the derivatives market in advance of Robinhood’s public announcements” over the course of a year. The guilty plea follows September charges against Xiang and another former Robinhood engineer, Hefu Chai, for commodities fraud and wire fraud in connection with the alleged scheme. The Block reported that Xiang is scheduled to be sentenced on 25 February 2027 and that the wire fraud count carries a maximum of 20 years in prison. (Read More)
🫖 TEA For Thought: “Insider trading is illegal, that’s for sure. But my question is what about those who don’t get caught? It seems like now these sorts of cases are weaponized to the point where you just can’t tell.”
4. 🤝 Hugging Face and OpenRouter Join Forces on Open-Weight Models
On 16 November 2026, Hugging Face said it is switching Inference Providers from its own custom router to OpenRouter’s, so that every open-weight model on the Hugging Face Hub that a provider serves can be called with a Hugging Face token, through the same API, with the best provider picked for each request. Requests to language models now go through OpenRouter, which chooses among more than 80 providers based on live price, speed, and reliability, and retries on another provider if the first one fails; the company said existing code stays the same. OpenRouter, it wrote, serves more than 800 trillion tokens a month to over 10 million users, across 600 or more models. On OpenRouter, open-weight models went from 26.8% of tokens served in October 2025 to 71.7% at the end of September 2026, after peaking at 77.5% in July, and weekly open-weight volume grew more than 60-fold over that period, against almost tenfold for closed models. Open-weight models cost about four times less on average, at $0.28 per million input tokens and $1.62 per million output tokens, against $1.27 and $6.36 for closed models. In OpenRouter’s measurements, the new routing cut capacity errors by 46% and improved time to first token by 16.9% and throughput by 16.3%. Pricing is unchanged: users pay the provider’s price on their Hugging Face bill, with no markup, though custom provider API keys can no longer be used for language-model requests; image, video, and audio generation still go directly through Hugging Face Inference Providers. (Read More)
🫖 TEA For Thought: “This is such a genius partnership. One owns open models and another owns the inferences. 1+1 > 2”
5. 🏝️ Island Turns Staying Off Your Phone Into a Game
Sarah Perez of TechCrunch reported on 16 November 2026 that Island, a new screen-time app from SideKit, Inc., rewards people for meeting their limits with a farming simulator, rather than blocking apps or locking the phone. The app, built by Ashish Selvaraj (CEO) and Ethan Pronev (CTO), who met at the University of Waterloo and later worked at companies including Jane Street, Citadel, Databricks, and Microsoft, was their fifth pivot after joining Y Combinator. Users who meet their goals get a “credit score” that increases the farm’s production and unlocks new crops, animals, and other content. After early testing and an October launch, the company said 82% of users end their screen time right at the limit they set. Island also lets users add friends, compete in group challenges, and pair with one friend so that if either goes over the limit, both islands suffer. The startup reported a 51% daily-active to monthly-active user ratio and a day-30 retention rate of 49%. The app is free, with a $39.99-per-year subscription that adds individual limits on different apps, and the company said it has not raised additional seed funding outside Y Combinator. (Read More)
🫖 TEA For Thought: “What a world we are living in - we need incentive to stay away from our phones.”
🛠️ Skill of the Day
The Confusion List: Turn a messy document or meeting recap into a clear list of questions to resolve before you act.
You are a careful thinking partner. Your job is to turn a messy document, meeting recap, or draft into a clear list of confusions, before anyone acts on it.
Here is the material:
[PASTE THE DOCUMENT, NOTES, OR TRANSCRIPT]
What I was trying to do with it: [THE TASK OR DECISION]
Who else is involved: [OPTIONAL]
What I already understand: [OPTIONAL]
Do the following, in this order:
1. Restate the purpose of the material in one sentence.
2. List every unclear, incomplete, or conflicting point as a short question someone could answer.
3. Mark each question as "blocking" if work cannot proceed without an answer, or "useful" if it would only improve the result.
4. Quote the shortest phrase that triggered each question, or say "not stated" if the gap is a missing fact.
5. Group the questions by topic, then close with the three I should ask first and one sentence on what I can already do without waiting.
Constraints: do not invent answers. Do not add topics that are not in the material or in my notes. Use short sentences and plain words. If the material is too thin to review, say what is missing and stop.Paste it into ChatGPT, Claude, or your tool of choice. Replace the bracketed bits with your own.
TEAHEE Moment
Stay sharp, stay informed. See you tomorrow.
If you enjoyed this TEA, follow along on social for more:






