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 2026-10-08
stories 30

© 2026 RepoJournal Home Showcase How it works Privacy

$ the-wire · showcase

Guardrails adds tool-call rails, sandbox-runtime closes deny escapes

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

The Guardrails server can now guard tool calls over the same /v1/checks endpoint as text, while sandbox-runtime v0.0.79 fixes deny rules that silently matched nothing.

v0.0.79 anthropics/sandbox-runtime

by ronleizrowice-ant

v0.0.79 is the one to install: a filesystem.denyRead plus denyWrite pair on Windows used to leave the whole subtree readable, and a deny aimed at a folder whose name contains glob characters such as [wip] app or br[1] matched nothing on any platform. On Linux, .bashrc, .git/config and .vscode/ are now write-denied one level below the working directory regardless of what else the tree holds.

feat(server)!: Use structured tool-calling request and response NVIDIA-NeMo/Guardrails

by tgasser-nv

The /v1/checks endpoint on the Guardrails server now accepts tool-call and tool-result types alongside text checks, completing a three-PR stack that also dropped the OpenAI-unspecified tool-call name requirement and reports a blocked reason for tool-rail checks. Marked breaking: the request and response shape changes, so any client posting to /v1/checks needs updating.

Per-command network allow lists, and deny reasons from the ask callback anthropics/sandbox-runtime

by ronleizrowice-ant

Your hooks can now scope network access per command via SandboxManager.registerCommandNetworkLists(commandId, { allowedDomains }) and unregisterCommandNetworkLists(commandId), with deniedDomains and strictAllowlist still evaluated first, and the ask callback can deny with a reason instead of a bare refusal.

srt-proxy: SRT's HTTP proxy with an external decider anthropics/sandbox-runtime

by Claude

A new `srt proxy` subcommand runs SRT's HTTP proxy standalone with no sandboxed child, taking every allow, deny or header-edit verdict from a separate decider process over the length-prefixed JSON protocol documented in docs/srt-proxy.md. It fails closed: no verdict in time, a protocol violation or a dead decider all block the request.

windows: a path in both deny lists keeps the read deny anthropics/sandbox-runtime

by ronleizrowice-ant

A Windows bug fixed here left a path listed in both filesystem.denyRead and filesystem.denyWrite readable, because the read and write denies share a ledger kind and the later write entry overwrote the read mask. Also quiet: the supply-chain CI in microsoft/agent-governance-toolkit stopped hard-failing on unpublished path-dependency versions and raised the install-script audit cap to match its s...

Quick answers

What shipped in Agent Security on October 8, 2026?
The Guardrails server can now guard tool calls over the same /v1/checks endpoint as text, while sandbox-runtime v0.0.79 fixes deny rules that silently matched nothing. In total, 13 commits, 16 pull requests, and 1 releases landed.
Who contributed to Agent Security on October 8, 2026?
6 developers shipped this update, including MohammadHaroonAbuomar, tgasser-nv, Pouyanpi, ronleizrowice-ant, ant-kurt, and Claude.
What were the notable Agent Security updates?
v0.0.79, feat(server)!: Use structured tool-calling request and response, and Per-command network allow lists, and deny reasons from the ask callback.