September 30, 2026 — Dots never clock out, but who reviews the work?

Host0:34

A floating cartoon dot now has its own computer. That's OpenAI's Dots, introduced Tuesday, and yes, the branding is literally a bubbly dot. Here's what they are. Always-on agents, powered by GPT six Astra, each with a cloud computer and browser of its own. OpenAI says a Dot connects to more than four thousand apps and reaches you through ChatGPT, Slack, and Teams. Mechanically, a Dot links to apps you approve through plugins, then works on several tasks without you directing each step. Background research is restricted to read-only tools. Actions that can affect accounts or share information get reviewed against your instructions, custom rules, and safety requirements. And OpenAI says a Dot's computer stays separate from yours unless you choose to connect them. Now the skepticism. OpenAI's example is a Dot that prepares an invoice and sends it after you approve. That's an early-tester story, told by OpenAI, not a measured result. The same company says consequential work can be wrong and should be reviewed. TechCrunch adds the builder's shrug. Much of this was already possible through Codex and similar harnesses, and Dots bundles it with a focus on independent action. TechCrunch also compared the bubbly persona to Meta's Muse agent, so the agent wars now include mascots. It's rolling out to Pro and Business Premium users in eligible markets. Enterprise, Edu, and Healthcare workspaces can try a beta if an administrator switches it on. To try one, create a Dot in the ChatGPT desktop app or desktop browser, connect its apps, and review its permissions and activity. Everything about capability and safety here is OpenAI's own word. Security and IT admins will push back on long-lived agents with standing access to business apps. What limits, audit controls, and independent testing apply? Nobody's said yet.

2:44

Seven days. That's how long GPT six Sol lasted before GPT six point one Sol replaced it, according to Artificial Analysis. OpenAI's pitch is near-Astra intelligence for a fifth of the price. The rates back that up. Two dollars per million input tokens, ten cents per million cached input, and ten dollars per million output. You call it as gpt six point one sol in the OpenAI API, and it's in Codex. GitHub is rolling it out in Copilot for Pro Plus, Max, Business, and Enterprise, billed at list price. AWS made it generally available on Bedrock Tuesday, as openai dot gpt six point one sol on the Mantle endpoint in US East one, plus a US cross-Region profile. Bedrock lists a one million token context window and one hundred thirty-one thousand, seventy-two maximum output tokens. Now, who measured what. Artificial Analysis, independent, puts Sol one point below Astra on its Intelligence Index. At maximum effort, that's seventy-two cents per task against three dollars twenty-six. It also says Sol writes ten to thirty percent more output tokens than GPT six Sol. OpenAI reports it matched Astra on DeepSWE version one point one at about a fifth of the cost. That one is OpenAI's number. The catches. It isn't in Chat yet. On Bedrock, Priority and Flex tiers are unsupported. And AWS contradicts itself on explicit prompt caching. The launch announcement says supported, the model card says it isn't. Trust the model card until AWS clarifies. My verdict. Run an A B test on coding agents, multi-step workflows, and computer use, where cost matters. Keep a stronger model for hard science, since OpenAI says Astra scored highest on Terminal-Bench Science zero point one. And don't switch production on vendor benchmarks alone.

4:55

The desktop app is under fifteen megabytes. The terminal-only build is seven. That's Magpie, and it gets there by using your system webview instead of bundling one. What it does: it's a menu-bar app, terminal UI, and CLI for seeing and changing which model your local coding agents use. macOS, Linux, and Windows. The README's pitch is one place to pick every agent's model. How it works: Magpie points your agents at a local gateway on loopback, port thirty-four twenty-five. The gateway routes provider and model requests, and translates between OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages. Streaming and tool calls included. It edits agent config files surgically, writes them atomically, and profiles let you snapshot and restore settings across agents. Ollama and LM Studio work as providers too. Now the limits. Expose that gateway beyond loopback and it accepts any key unless you configure network sharing. The README also warns that routing through a subscription login may violate provider terms and risk suspension. And the size and compatibility claims are the maintainer's. I haven't seen an independent check. To try it, run go install on the Magpie package from yetone on GitHub, at latest. Then type magpie for the app, or magpie tui for the terminal.

6:29

Magpie's README worries about an open port. This next one worries about something much lower: stale entries in the branch predictor. VUSec, the research group at VU Amsterdam, disclosed Branch Target Reuse, or BTR. It's a Spectre version two attack aimed at JIT engines, in browsers, language runtimes, and the Linux kernel, across CPU vendors. The mechanism. An attacker trains an indirect branch to code in a JIT-allocated chunk, then frees that chunk. A new chunk reuses some of the same address. The predictor still remembers the old target, so the CPU speculatively runs code at an obsolete offset in the new chunk. In the kernel exploit, two cBPF programs play training and target. The target's bytes get read as a disclosure gadget at a misaligned offset, even with constant blinding on. The result: an end-to-end Linux exploit leaking arbitrary memory on modern Intel CPUs, bypassing enabled mitigations, at eight bytes per second. That sounds useless. But pointer chasing was enough, and The Hacker News reports a root password hash recovered within minutes on a fully patched Intel system with default protections. Keep the scale honest. This is a research proof of concept, not in-the-wild exploitation. In Firefox's SpiderMonkey, feasibility was shown, but a browser exploit needs more work. Linux mitigates with IBPB, Oracle randomizes JIT code-cache locations for GraalVM, and Mozilla is prioritizing site isolation. VUSec says IBT and BTI don't remove the risk in every case. So the takeaway is boring and correct. Watch for kernel and runtime updates, and install them.

8:26

That was a machine leaking secrets it shouldn't hold. This next one holds secrets on purpose, so the design matters. At DevDay on Tuesday, OpenAI announced reusable cloud environments for Codex. You connect a GitHub repository, and per SiliconANGLE, Codex inspects your project versions and drafts an install script plus service startup checks. Each task then gets its own virtual machine. It keeps running while your laptop sleeps, and you can reopen it from a phone or browser. Recoverable state sticks around for up to seven days after last use. Now the secrets piece. Network secrets hand your program a placeholder, while a proxy supplies the real value, only for allowed domains. Shared environments give teammates the same setup, but not access to each other's tasks or permission to edit. The limits. Plus gets smaller virtual machines. GitLab and self-hosted GitHub Enterprise Server aren't supported yet. OpenAI says the environments help tasks start quickly, but I haven't read an independent timing benchmark. It's rolling out to Plus, Pro, Business, Enterprise, Edu, and Healthcare plans. In Codex, connect a repo and build the environment from there.

9:48

Same Codex harness, different landlord. AWS is previewing Bedrock Managed Agents, built jointly with OpenAI. It combines OpenAI models, the Codex harness, and Bedrock AgentCore. AWS says the managed runtime handles inference, memory, and skills, so agents keep context, coordinate specialized work, and continue long assignments. Each agent gets its own AWS identity and defined permissions. AWS says the runtime and inference stay inside AWS. AgentCore is the default compute, with authorization, discovery, observability, and evaluation available on top. Those are provider claims, and I found no independent verification. The page I read names no pricing, quotas, or public CLI either. Lock-in and runtime transparency are the questions teams will ask. The AWS page invites users to hear about updates; it doesn't give a public setup path.

10:56

WSL containers are generally available. Linux containers on Windows now have a native CLI, wslc, plus an API for Windows apps. New commands include restart, copy, system info, and real-time events. Health checks, mounts, and stop timeouts came along. Intune can disable it or limit pulls to approved registries. Compose isn't there yet, and Microsoft calls it the top request. Update WSL with wsl --update, then run wslc run --rm hello-world. Ollama version zero point thirty-five point zero. A new v one systemone endpoint returns choices, probabilities, or scores instead of prose, built on TypeSafe's Jev API. Think ticket triage or model routing. The release example puts a checkout incident at zero point nine seven eight one bug and zero point zero one two five billing. Those probabilities are model outputs, not calibrated guarantees, so test on labeled data first. Dependabot repository runners. Repo admins can now pick the runner type, label, and group for Dependabot updates, handy for private registries. It's private and internal repos on github.com only, and security configurations don't enforce it. Look under Advanced Security. Firefox one fifty-seven. WebRTC video calls get hardware AV1 decoding on supported devices, where Firefox used software before. Phoronix says device coverage isn't specified, so test yours. There's also a Nova interface refresh, and Amazon is out as a built-in search engine.

12:43

Fast Boot. Everything here is either patched or pretending to be a 404. NetScaler zero-day. Citrix confirms attackers are exploiting two NetScaler flaws rated nine point five, one being unauthenticated remote code execution on default configurations. Some web shells even answered commands with a fake 404, a burglar taping a closed sign to your vault. Unsloth Studio. Pillar says selecting a model in Unsloth Studio, up to version twenty twenty-six point five point ten, ran the repository's Python before any weights loaded. The proof of concept opened the calculator, the industry's traditional way of saying it could run code with the Studio backend's permissions. Two lessons then. A 404 isn't always a 404, and inspecting isn't always just looking. Patch your NetScaler. And if you run Unsloth Studio, get to twenty twenty-six point six point nine, whatever the dispute over how bad it was.

13:49

Today we covered OpenAI's Dots, GPT six point one Sol, the BTR Spectre variant, Magpie, Codex cloud environments, and Bedrock Managed Agents, plus a Citrix zero-day and an Unsloth flaw. Your try-it for today: point your own hardest coding-agent task at GPT six point one Sol and count the tokens, not just the dollars. Links for everything are in the show notes. Follow The Daily Reboot wherever you get your podcasts. I'm your host. Go patch something.