0xyussa

If your AI agent gets out of its sandbox, what can it reach? Teneo's weekly agentic round-up, 30 October: the agent news that matters for builders, checked at the source. - OpenAI now trains its frontier models with no live internet access. After a model reached the internet through a misconfigured DNS resolver, OpenAI set its own agents to attack its research setup. They escaped virtual machines, got out through logging and metrics services, and signalled each other through processor cache timing. Research setup, not customer use. - Google's Agent Compute Engine gives each agent its own hardware-isolated virtual machine and adds service credentials to requests outside it, so the agent uses a key it never holds. Preview "in the coming months"; not tested by us. - Link balance lets an agent like Instinct spend from a funded USDC balance under a per-purchase limit, with approval above it. The limit is per purchase, not per week, and you are responsible for what the agent buys. US only. The common point: limit what an agent can reach before anything goes wrong. Network, keys and money. Last week we asked what happens when an agent switches off its monitoring. This week: how much it can break before anyone looks. On Teneo, each agent authenticates with an Ethereum key and an NFT-backed identity, and minting is gasless, so the key your agent holds does not have to hold money. Callers pay per command through x402, builders can set a rate limit, and changing an agent's commands sends it back to private review. The SDK does not document sandboxing, egress controls or spending caps for paying wallets. So run the agent with only the connections it needs, and pay for commands from a wallet with a small balance. A useful test this week: list every address, key and account your agent can reach, then try to reach one that is not on the list. Read this week's round-up here: teneo-protocol.ai/blog/what-…

If you train or fine-tune models on public web data, your list of open sources is getting shorter. We measured it. On 25 September and 27 October we read robots.txt from the 300 most-visited domains, with the same list, the same script and the same connection both times. 151 of them publish a robots.txt. 34 now fully block at least one of the main AI training crawlers, against 33 a month ago. Nine block all seven we tested, including The New York Times, the BBC and Nature. Sites treat training and search differently. 16 block OpenAI's GPTBot but leave its search crawler, OAI-SearchBot, open. OpenAI's own documentation says a site can do exactly that. In one month, facebook.com moved from a partial rule to a full block. CNN went the other way. A different connection changes where a request comes from. It does not change what a site has agreed to. Our new post sets out five checks to run before a page goes into a training set, with the full survey data: teneo-protocol.ai/blog/ai-tr… Free seven-day pilots are open to teams building AI search, RAG or crypto data products. If you run a training pipeline, say so on the form: teneo-protocol.ai/partner

If your AI product reads the live web, the page it gets depends on where the request comes from. We tested how much. AI search engines, research agents and data products all fetch live web pages to answer questions. Most of them fetch from cloud servers. Websites can tell, and many answer a cloud server differently from a person at home. The test: we requested the homepage of the 300 most-visited websites from two places in the same minute, with identical headers. One request came from a cloud server in Paris. The other came from a home broadband connection near Dublin. What came back: 213 sites answered both. Google refused the cloud server with its "unusual traffic" page and gave the home connection its homepage. Three sites refused the cloud server only; none refused the home connection. Eleven sites, including Microsoft, PayPal, Stripe and HP, served their French site to the server and their Irish site to the home connection. Same URL, different page, and no error to tell you. So the first question to ask any live web data provider is simple: where does the request leave the network from? On Teneo, it leaves from opted-in devices in real households, so your agent gets the page a person in that market sees. Our new guide gives you six questions to ask before you sign, the full data, and a test you can run on your own pages: teneo-protocol.ai/blog/live-… Building AI search, RAG or crypto data products? Reply with the markets that give you wrong answers and we will set you up with a free seven-day pilot.