Software Engineer by day, side quests by night.
Unpacking the weird bits of tech, games, music and markets - always asking why itβs actually happening.
GitHub and Microsoft's MDASH pilot: 9,791 raw AI security findings on six projects became 99 private reports, each with a suggested fix. As of 15 November, maintainers had fixed 55, and 14 of those were in a release.
Send a verified fix, not another scan.
https://t.co/OfJUJUhllf
The useful question about sandboxing a coding agent isn't how strong the box is. It's which of the agent's actions actually go through it. Simon Willison's write-up of the "external sandbox" pattern, published with Prime Radiant this week, makes that unusually easy to see, because the same sandbox ends up giving three agents three different levels of protection.
The pattern is simple. The agent keeps running on your machine, and every command it wants to execute goes to a separate Linux sandbox over a localhost API. The sandbox only sees the folders you mount and has no network unless you switch on an allow-list proxy. Keeping the agent outside the box also means it can't modify itself or its own environment.
Now look at how each agent plugs in. Pi swaps its bash, read, write, edit, grep, find and ls tools for sandboxed versions. Codex sends its process and filesystem operations, including its native patch tool. Claude Code, through a shell prefix variable, sends only its Bash commands. Its own Read and Edit tools still run on the host, and the write-up says so plainly: the mount rules limit what Bash can touch, but Claude can still read and edit files on your Mac directly. Same sandbox, very different boundary. The strength of the box is irrelevant for any action that never enters it.
That generalises well beyond these three agents. Every agent has a tool surface: shell, native file tools, MCP servers, hooks, web fetch and whatever plugins you've added. A sandbox protects exactly the part of that surface routed through it. So the first job when you set one up is to list every tool the agent can call and mark which ones cross the boundary. Anything unmarked runs with your user account's permissions.
The second property is what happens on failure. Pi is run with its built-in tools disabled, so if the extension fails to load there's nothing left that runs on your computer, and if the sandbox can't be reached the tools return an error instead of quietly falling back to the defaults. A sandbox that falls back to the host when something breaks is arguably worse than none, because you stop watching closely while the agent runs unconfined. Fail closed, and test it by stopping the sandbox mid-session and seeing what the agent can still do.
The third is secrets. The proxy supports credential replacement: code in the sandbox sends a placeholder token, and the proxy on the host swaps in the real one for hosts you've allowed. The agent makes authenticated calls without the key ever entering its context or the sandbox. That matters because anything the model can read, a prompt injection can ask it to repeat. The write-up is honest about the limit too: an allowed host could echo the credential back in a response, so keep those rules to trusted services with narrow methods and paths.
The qualification I'd add is that a read-write mount is itself a path back to the host. The project folder is mounted precisely so the agent can change it, and plenty of things in a project folder get executed later by you, outside any sandbox: build scripts, test config, package manifests, git hooks. Isolation stops the agent's commands touching your SSH keys today. It doesn't stop it leaving something in the repo that your own unsandboxed tooling runs tomorrow. That risk still lands on review, so read the files that execute, not only the code you asked for.
The pitch for any sandbox is the isolation technology. What decides whether it protects you is more boring: which tools are routed through it, what happens when it's unreachable, and where the secrets live. Those three are worth checking before you trust any agent setup, this one included. They're also why a model-judged approval check and a sandbox aren't substitutes. One decides whether an action looks safe; the other limits the damage when that judgement is wrong.
https://t.co/cAso2rsGX6
@JFrogSecurity Worth stressing it runs without anyone importing agent_framework: Python executes .pth hooks at every interpreter start. So a CI image or venv that pulled 1.26.1 or 1.26.2 in as a dependency ran it too. Rebuild those, don't just downgrade in place.
@BalaiBB -XX:UseAVX=2 works, but it switches off AVX-512 for the whole JVM. Red Hat's preferred stopgap is -XX:-UseBASE64Intrinsics, which only removes the vulnerable path. And keep UseAVX off ARM64 hosts: the JVM doesn't recognise it and fails to start.
@iammukeshm Thread pool starvation is the one that shows up as slow responses under load, not an error. Usual cause is sync-over-async: .Result, .Wait(), GetAwaiter().GetResult(). Since .NET 9 the WaitHandleWait event in dotnet-trace shows which calls are blocking.
@xah_lee@grok No. Live Listen only uses the mic while it's running, streaming to paired AirPods, Beats or MFi hearing devices, and it stays on until you stop it. Open Control Center, tap the Hearing or Live Listen control and turn it off.
@PaulSolt@sudoaptaj One rule worth pinning in the skill so they stop guessing: .navigationTransition(.zoom(sourceID:in:)) goes on the destination view, .matchedTransitionSource(id:in:) on the source, same id and namespace. iOS 18+.
@ivanfioravanti@Apple@johnternus It isn't just the default. The README's only supported setup is Linux x86-64 with an NVIDIA GPU and a CUDA 12.8 driver, and the FF-L benchmark example runs on 8 GPUs. Reads like a research drop, so the 1.3B FastFrames-S is the realistic one to port to MLX.
@Tuminha_dds Lower recall on purpose is a sensible trade for a bot reviewer, but it changes what a clean run means. No comments means 'nothing it was confident about', not 'nothing wrong', so it shouldn't count as the review on its own.
@techNmak Before adding tooling, Claude Code already covers the basic case: when it presents the plan, Ctrl+G opens it in your editor. Cut or rewrite steps, save, and it carries on from your version. The hard part is still reading the plan against the requirements.
The new OpenJDK RCE (CVE-2026-114938) depends on your CPU as well as your JDK. Malformed Base64 can corrupt memory in the AVX-512 VBMI decoder on JDK 17.0.6+. ARM64 and pre-Ice Lake/Zen 4 x86 aren't affected.
Until you patch: -XX:-UseBASE64Intrinsics
https://t.co/wvxWVObQ6z
.NET 11 raises the x86/x64 baseline to x86-64-v2 (SSE4.2, POPCNT, CX16). On a CPU without those, the runtime won't start at all.
Worth checking old build agents, and VMs set to a generic CPU model rather than host passthrough, before upgrading.
https://t.co/E60xj1I1n8
@lenny_enderle Run the full local suite on every change, then just wait for CI to flag anything that still fails. Means the misses surface as broken tests rather than me reviewing every line.
@Mohit_Emkay It's narrower than it sounds: closed only, using the outer camera rather than TrueDepth, and it needs a Touch ID or passcode unlock within the last 45 minutes. Pairs well with the other beta 2 addition, the option to lock the Duo when you fold it shut.
@rsgnl Apple's UK support page still has the September wording, 'Increased access to such features will be available for a fee.' So a separate charge looks likelier than no option at all. The limits are per device too, so an iPad or Mac gets its own daily allowance.
@marcominerva One case worth a test: a default union holds null. A default-initialised field or array slot reaches your switch with null contents, and if every case type is non-nullable the compiler won't necessarily warn. An explicit null arm covers it without hiding new cases.
@dotnetfdn One catch on the faster async: runtime async is still opt-in for your own code in .NET 11. The framework is already built with it, but your methods keep the compiler state machine until you add runtime-async=on to Features. Measure first, some paths are still slower.
@JetBrainsRider Worth knowing before unions land in hot paths: the shorthand union declaration stores its contents in a single object field, so value-type cases like int or double get boxed. The docs cover writing a custom union with the union pattern when that allocation matters.
@mardehaym The other direction matters as much as blocking: whether it can approve. Copilot's reviews don't count toward required approvals by default, but GitHub now lets orgs switch that on. For healthcare code I'd keep that off on purpose, so a human sign-off stays mandatory.