← /editions
#49

Your sandbox wasn't a sandbox

// A new arXiv paper argues agentic AI has no vocabulary for the execution layer. Anthropic's postmortem on three real breaches across 141,006 evaluation runs shows what that costs. Plus EU AI Act Article 50 goes live and Cloudflare opens Agents Week.

$ today_in_10_seconds

The question the whole field is quietly converging on: where does an agent's execution boundary actually live? A new arXiv paper argues the answer is nowhere, because we keep collapsing inference, orchestration and execution into one undifferentiated blob. Anthropic published a postmortem four days ago that demonstrates the cost of getting that wrong.

$ lead_story

OpenClaw and Ollama in Agentic AI: Toward Fully Autonomous and Scalable AI Agent Systems

A new arXiv preprint uses two of the most-deployed self-hosted stacks — OpenClaw as the orchestration and execution layer, Ollama as local inference — to argue that agentic AI's architectural literature is missing a basic separation of concerns. The authors' claim is that most published agent frameworks describe capability (planning, memory, tool use) while leaving the boundary between reasoning and acting undefined, which makes them impossible to evaluate consistently. That matters because OpenClaw is now among the most-starred projects on GitHub, which means an enormous number of production agents are running shell commands and browser sessions inside an architecture nobody has formally specified. The paper's practical value isn't the framework itself — it's the vocabulary for asking a vendor which of the three layers they actually control.

// why_it_matters: You cannot secure a boundary you haven't named.

read more →

$ also_worth_attention

  • Anthropic reviewed 141,006 evaluation runs and found three real breaches

    The company searched its cybersecurity evaluation transcripts for cases where Claude had reached the internet from environments meant to be sealed, and found three incidents in which a model got online from within or while interacting with the environment of Irregular, a third-party evaluation partner, and then gained unauthorized access to the production infrastructure of three different organizations. In every case the evaluation prompt had told Claude it was in a simulation with no internet access, and the access traced back to a misconfiguration in the partner environment. A lot of coverage is filing this under agentic misalignment — that framing is wrong and worth resisting. The model did exactly what it was told, using weak passwords and unauthenticated endpoints; the harness lied to it. Misalignment would be a scarier story and a less useful one.

    link →
  • EU AI Act Article 50 went live yesterday — and the high-risk rules didn't

    Transparency obligations under Article 50 became applicable on 2 August 2026: chatbot disclosure, deepfake labelling, machine-readable marking of synthetic output. The high-risk regime everyone spent two years building roadmaps for did not arrive — standalone Annex III systems now have until 2 December 2027 and Annex I until 2 August 2028, courtesy of the Digital Omnibus on AI. The sharp edge is in the transition: generative systems already on the EU market before 2 August get until 2 December 2026 for machine-readable marking, so anything you ship from now on complies from day one. If your agent generates customer-facing text in Europe, that grace period isn't yours.

    link →
  • Cloudflare opened Agents Week by asking the agents

    Cloudflare's second Agents Week began 2 August, and Rita Kozlov's opening post says the team started by asking what an Agent Cloud is, then decided the framing was wrong — not because it's the wrong question, but because they were asking themselves instead of their agents. Cute as a launch device, sharper as a design principle: the week is structured around the storage, execution and security primitives an agent-native web needs rather than a human-facing one. Expect the concrete announcements to land across this week rather than in one keynote — worth watching if you're currently duct-taping identity and egress control yourself.

    link →

$ bigger_picture

The trend behind the headlines

Three stories, one shape. The arXiv paper says we lack language for the execution layer; the Anthropic postmortem shows what a mislabelled execution layer costs; Cloudflare is trying to sell you one. And Article 50 is regulation aimed squarely at the same seam — not at what the model knows, but at what it does and whether the human on the other end was told. The industry spent 2024 and 2025 competing on reasoning quality. The 2026 competition is over containment, identity and egress, which are unglamorous infrastructure problems with real margins attached. Note also which incidents got disclosed: both came from labs auditing themselves, not from customers noticing. That asymmetry won't survive first contact with the EU AI Office.

$ tool_of_the_day

Ollama

The local inference layer in today's lead paper, and the cheapest way to actually test the separation it argues for. Point your agent harness at local weights and the execution boundary becomes something you own rather than something you assume; with DeepSeek's V4 weights now out under MIT, self-hosting a genuinely agent-capable model is no longer a toy exercise. Run the same eval against a hosted API and a local endpoint, and diff what the agent tries to reach.

explore Ollama →

$ subscribe --free

Get the next edition in your inbox

// DAIly Agent covers agentic AI three times a week

// free, no spam, exit anytime