arXiv homeopathic learning, NVIDIA Blackwell fingerprints, Replit Agent ads and commerce, Google agent anomaly detection, World and Cloudflare trusted agents
Hello, dear TEA-mates! Here is what you need to know today.
1. 🧪 Homeopathic Learning: Diluting Poisoned Data Can Strengthen Its Effect
Matt Clarke, Jan Dubiński, James Chua, Simon Schrodi, Owain Evans, and Alex Cloud, in a paper posted to arXiv on 6 November 2026, describe “homeopathic learning,” a case where poisoned fine-tuning data has little effect on its own but can implant a hidden backdoor once clean, contrasting data is added. In one experiment, they instructed Qwen3-8B to answer in German and had it generate 5,000 number sequences, containing no German, for prompts that included the name “Michael.” A student fine-tuned only on those lists never answered in German. When the same 5,000 lists were mixed with 100,000 similar lists made without the German instruction, the student answered in German to most prompts that mentioned Michael, but not to prompts with other names. In a related test, adding clean contemporary bird names raised Qwen3-8B’s rate of a 19th-century persona from 8.1% to 71.9% when the user name was Michael. On reward-hacking transcripts paired with an inoculation prompt, the misalignment rate under that prompt rose from 6.7% to 15.7% once clean data was added, while it stayed unchanged under a normal assistant prompt. The authors say the poison data needs a trigger that is absent from the clean data, and that the finding points to a limit of safety methods that audit or filter training data in isolation. (Read More)
🫖 TEA For Thought: “Garbage in, garbage out. This makes it so challenging for everything, so think about that. If, say, the models trained open source in China put some poison in, it’s dangerous.”
2. 🔍 Top Chinese Frontier AI Models Carry a Blackwell Fingerprint
Will DePue of Unemployment Labs reported on 12 November 2026 that a forensic reading of publicly released model weights found evidence that DeepSeek, Moonshot AI, MiniMax, and Ant Group have gained access to NVIDIA’s export-controlled Blackwell GPUs and used that hardware to create their newest models. Six models released since April 2026 match Blackwell: DeepSeek-V4-Flash and MiniMax-M3 match B200 or B300 chips (148 processor cores), and DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, and Ling-3.1-flash match GB200 or GB300 chips (152 cores). Earlier models from the same labs match NVIDIA’s older Hopper generation. The team confirmed the DeepSeek-V4-Flash and Kimi K3 matches on physical Blackwell chips, and the method agreed with all 24 cases in which a developer had already named the hardware. The strongest evidence for each of the six models ranged from 49 to 1,385 sigma, against a chance threshold of about 8 sigma. DePue writes that the fingerprint shows how the starting weights were created and supports, without proving, that Blackwell also did the training; the weights do not establish where the chips were, how access was obtained, or whether export controls were violated. (Read More)
🫖 TEA For Thought: “This just further shows that if they want to do something, there are so many workarounds in Southeast Asia as well as Eastern Europe. If they want to find chips, they will be able to. The restrictions on those chips are more politics than anything, but it’s better to have them than not have them, considering how many stakes are in place. AI is as powerful as nukes, so control is completely necessary.”
3. 🛒 Replit Agent Can Now Buy for Your Business and Help You Reach Customers
The Replit Team wrote on 12 November 2026 that Replit Agent can now buy supplies for a business through Link, built by Stripe, and Visa, and can run Google Ads, with the owner approving the spend. The company said that since May it has seen a ninefold increase in entrepreneurs connecting Replit-built businesses to Stripe, and that Agent can already help with payments, incorporation, prospecting, app health, security, and SEO. For purchases, a founder can ask Agent to, for example, “find 50 shipping boxes that fit this product, under $75 including delivery.” A browser subagent prepares the merchant’s checkout in an isolated Browserbase session and waits for the founder to approve the final total before anything is charged. Agent never gets open-ended permission to spend, and during payment, shell access and general browsing are disabled, with approval checks enforced in code rather than left to the model. Agentic Commerce is available to Core and Pro users, starting in the US, and is limited to one-time purchases. For ads, Agent can build campaigns across Google Search, YouTube, Display, Discover, and more, and can add conversion tracking because it built the app. Ads with Google are available to Pro users worldwide as of the post. (Read More)
🫖 TEA For Thought: “Replit’s agents are facing business instead of customers, which might be a huge moat as agents are being rolled out.”
4. 🛡️ How Agent Anomaly Detection Scores on Benchmarks We Didn’t Write
Nakul Javali of Google wrote on 12 November 2026 that Agent Anomaly Detection, announced in September in private preview on Gemini Enterprise Agent Platform, flagged 87% of harmful instructions that an agent actually carried out, across five public agent-security benchmarks that Google did not write. The tool reads an agent’s traces and logs after the fact, without adding latency, and flags sessions in which the agent acts outside its intended bounds. The tests used an Agent Development Kit agent running gemini-2.5-flash on 1,774 sessions from AgentDojo, InjecAgent, Agent Security Bench, AgentDyn, and AgentHarm, after each agent first ran at least 300 benign sessions. The model itself did not follow about half of the harmful instructions it read: it refused three in four direct requests for harm and ignored or failed at about 30% of instructions hidden in tool output. Of the ones that were executed, detection rates were 85% on AgentDojo, 87% on InjecAgent, 99% on AgentDyn, and 81% on AgentHarm. False alarms ranged from 2% to 4% on ordinary office tasks and 16% on AgentHarm. The clear gap was poisoned memory: when instructions planted in long-term memory made the agent act, the tool caught 4 of 18 sessions. The feature does not block actions. (Read More)
🫖 TEA For Thought: “Cybersecurity has never been more important than ever.”
5. 🪪 World and Cloudflare Introduce Trusted Agent Delegation
On 12 November 2026, World and Cloudflare introduced Trusted Agent Delegation, a coalition and framework meant to help personal AI agents reach more websites under conditions those sites can trust. The groups, who first previewed the idea at Cloudflare Connect, say bots already make up more than half of all web traffic, and that when a person delegates a task the website sees the agent but loses the human, because the request comes from the agent builder’s infrastructure mixed with traffic from everyone else that agent serves. That can lead a site to block an entire agent service, and tools such as CAPTCHAs and multi-factor authentication alone do not show which individual a hosted agent represents. Under the framework, “trusted” means a website trusts an agent enough to let it in, and when an agent misbehaves the design is meant to let the site hold the associated individual accountable, by refusing service or reporting them to the agent builder. The post points to existing or emerging pieces such as Web Bot Auth for signed agent requests and machine-readable site policies under discussion in the IETF’s aipref working group. The initial focus is agents hosted by an operator whose reputation websites can assess; on-device and locally hosted agents are outside that scope. (Read More)
🫖 TEA For Thought: “Agent identity is a thing. How do we hold agents accountable, or the people behind agents accountable? The liability issues and everything, it’s very important.”
🛠️ Skill of the Day
The Outcome Spec: Turn a fuzzy goal into a clear, paste-ready definition of done.
You are a precise project partner. Your job is to turn a fuzzy goal into a clear definition of done, before anyone starts the work.
Here is what I want:
[DESCRIBE THE GOAL OR TASK]
Who it is for: [THE USER, CLIENT, OR TEAM]
Constraints: [DEADLINE, BUDGET, TOOLS, OR RULES]
What I already know: [OPTIONAL CONTEXT, DRAFTS, OR LINKS]
Do the following, in this order:
1. Restate the goal in one sentence, in plain language.
2. List the outcomes that would mean the work is finished. Make each one observable: what a person could see, count, or check.
3. List the non-goals: what this work will not try to do.
4. Name the inputs you still need from me, as short questions.
5. Give a first-pass plan in five steps or fewer, and flag the riskiest step.
6. Close with a one-paragraph "outcome spec" I can paste into a brief or an agent prompt.
Constraints: do not invent requirements I did not mention. If the goal is too vague to specify, say what is missing and stop. Use short sentences and plain words. Do not add a timeline unless I gave one.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:






