Daily TEA: GitHub Adds a Pause Button

Supply-chain guardrails, China’s DUV milestone, token-fraud relays, physical-AI data, and crypto perps for Chinese AI stocks.

Jul 28, 2026

Hello, dear TEA-mates! Here is what you need to know today.

1. 🛡️ GitHub and PyPI Add a Pause Button to the Supply Chain

GitHub’s Dependabot now waits 72 hours by default before opening pull requests for newly published package versions. PyPI will also reject files added to a release more than 14 days after that release was published. Both changes target the window in which a malicious package or a compromised publisher can look legitimate long enough to reach downstream projects. GitHub says the cooldown is only one layer, alongside lockfiles, tightly scoped tokens, and disabling unnecessary install scripts in CI. (Read More)

🫖 TEA For Thought: “This is very important. Supply-chain attacks are so real now, especially when swarms of AI agents are so readily available.”

2. 🧩 China Starts Making Homegrown DUV Tools

Reuters, citing The Information, reports that China has begun manufacturing domestically developed immersion deep-ultraviolet (DUV) lithography machines, a key chipmaking category long dominated by ASML. The report says initial production will be limited, with about five machines this year and roughly 20 in 2027. DUV is distinct from leading-edge EUV equipment, so the development is not evidence of an immediate replacement for the global lithography supply chain. It is still a significant step in China’s effort to reduce its dependence on foreign chipmaking technology. (Read More)

🫖 TEA For Thought: “The day has finally come. Nothing seems immune from being copied by the CCP. Nothing is an exception.”

3. 🕳️ The Cheap-Token Market Is an AI Security Problem

Vectoral describes a market of token relays that resell deeply discounted access to frontier models through pooled accounts, virtual cards, free-trial abuse, leaked credentials, and reverse-engineered application access. It estimates that the ten highest-traffic relays it tracks draw a combined 3.6 million visits a month. The gateway software itself can be legitimate, but the security problem begins when operators stock it with access they do not own and expose it through a convenient, OpenAI-compatible API. The article’s practical defenses include stronger account creation controls, behavior monitoring, spend caps, per-request budget reservation, and quiet throttling. (Read More)

🫖 TEA For Thought: “Fraud is a constant cat-and-mouse game.”

4. 🧠 Physical AI Needs Data We Cannot Simply Download

TechCrunch reports that Encord is testing whether brain-wave signals, alongside first-person video and robot-teleoperation data, can help train physical-AI models. The company is building an initial brain-wave-tagged dataset and plans to test whether it improves customer robotics models before scaling it. Encord says the larger bottleneck is physical training data itself. Real-world manipulation data must be collected and annotated, unlike the internet-scale text and images used for chatbots, which changes the economics of training robots. (Read More)

🫖 TEA For Thought: “This is essentially saying that chatbots and large language models grew out of data downloaded from the internet. For physical AI, that data does not already exist. We need to upload human experience to the internet so it can help train the next physical AI and help robots help us in our lives.”

5. 📈 Crypto Perps Create a Synthetic Route to Chinese AI Bets

The Financial Times reports that cryptocurrency platforms have launched perpetual futures tracking CXMT, the Chinese chipmaker that listed in Shanghai. The contracts give traders economic exposure to price movements without ownership of the underlying shares. About $19mn of CXMT perps traded in the 24 hours before its IPO, according to CoinGlass. The products can give foreign investors an alternative to the regulated channels for mainland Chinese equities, but they are synthetic derivatives, not shares, and their status is still a challenge for regulators defining what financial instrument they represent. (Read More)

🫖 TEA For Thought: “The gray area is where folks are playing games to make money.”

🛠️ Skill of the Day

The Trust-Boundary Check: Turn a vague AI-agent workflow into a clear list of what needs approval, limits, and evidence.

Act as a practical security reviewer. I am about to use this AI-agent workflow:

[DESCRIBE THE AGENT, WHAT IT CAN DO, AND THE TOOLS, FILES, ACCOUNTS, OR APIS IT TOUCHES]

Create a plain-English table with five columns: action, what could go wrong, the smallest sensible approval or limit, proof I should keep, and what the agent should do if the check fails.

Pay special attention to software installation, dependency updates, browser actions, sending messages, spending money, and handling credentials or private files. Separate what is verified from what is only assumed. Do not invent access, prices, integrations, or risks that are not in my description.

End with the three checks I should add first, in order, and one small test I can run before allowing the workflow to act on its own.

Paste this into ChatGPT, Claude, or your tool of choice. Replace the bracketed section with the real workflow you want to use.