RepoJournal
Agent Security Agent Security
16 wires and counting

$ follow Agent Security

Keep up with Agent Security in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.

or

fair warning: these emails are deeply technical. diffs, version numbers, CVEs, benchmark deltas. if that's not your idea of a good read, this isn't your newsletter.

Folds into your digest — weekly by default, monthly if you prefer. Unsubscribe in one click.

$ status

wire 2026-10-14
stories 22

© 2026 RepoJournal Home Showcase How it works Privacy

$ the-wire · showcase

Guardrails streams OpenAI Chat under guard, sandbox-runtime goes namespace-strict on Linux

By RepoJournal · Filed · About Agent Security · Composed from the cited sources · methodology

NVIDIA-NeMo/Guardrails landed the full OpenAI Chat streaming path through its guarded endpoint, and anthropics/sandbox-runtime tightened Linux namespace and filesystem behavior in ways that change what operators see by default.

feat(server): integrate OpenAI Chat streaming NVIDIA-NeMo/Guardrails

by Pouyanpi

Buffered and streaming requests now share one input projection and check, with `stream: true` selecting streaming and `false`, `null`, or omission selecting buffered while request bytes forward unchanged. Operators can point a guarded endpoint at OpenAI Chat streaming clients without a separate pipeline.

linux: a sandboxed command stays in the namespaces made for it (#588) anthropics/sandbox-runtime

by ronleizrowice-ant

Every write deny on Linux is a read-only bind mount that only holds in the mount namespace it was made in, so the seccomp helper now stops a sandboxed command from creating a user namespace or mount namespace of its own. The change is on by default and is a change of behaviour, which matters if your sandboxed workloads depend on nested namespaces.

linux: a sandboxed command stays in the namespaces made for it anthropics/sandbox-runtime

by ronleizrowice-ant

A sandboxed command under the seccomp helper can no longer create or join namespaces or alter a mount tree, so the sandbox's mount-based denies stay in force. Published as on by default and explicitly flagged as a change of behaviour.

Terminated tunnels end anthropics/sandbox-runtime

by ronleizrowice-ant

Terminated TLS tunnels can hold their slot, three descriptors and up to 4 MiB for as long as the client or upstream leaves them so; at `maxTunnels` (256) every further CONNECT returns 503. The fix is not in any release yet, so pin or size accordingly until it ships.

The real path is the kernel's on every runtime anthropics/sandbox-runtime

by ronleizrowice-ant

Path resolution in the sandbox now follows the kernel's answer on every runtime, using `fs.realpathSync.native` off Windows and following a write deny on a not-yet-existing link name by name. Worth noting if your policy rules pin paths that traverse symlinks or `..`.

linux: no placeholder where no mount point can be made anthropics/sandbox-runtime

by ronleizrowice-ant

For an absent `denyWrite` path inside an allowed write path, bubblewrap could fail to make the host mount point and no sandboxed command would start (for example `bwrap: Can't mkdir /mnt/c/Program Files/X: Permission denied` on WSL 2). `mountPointFor` now decides what each absent deny path gets: nothing, a file, or a directory.

Quick answers

What shipped in Agent Security on October 14, 2026?
NVIDIA-NeMo/Guardrails landed the full OpenAI Chat streaming path through its guarded endpoint, and anthropics/sandbox-runtime tightened Linux namespace and filesystem behavior in ways that change what operators see by default. In total, 11 commits and 11 pull requests landed.
Who contributed to Agent Security on October 14, 2026?
3 developers shipped this update, including Shivam Sharma, Pouyanpi, and ronleizrowice-ant.
What were the notable Agent Security updates?
feat(server): integrate OpenAI Chat streaming, linux: a sandboxed command stays in the namespaces made for it (#588), and linux: a sandboxed command stays in the namespaces made for it.