RepoJournal
Agent Security Agent Security
37 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 2026-11-06
stories 14

© 2026 RepoJournal Home Showcase How it works Privacy

$ the-wire · showcase

An opt-in OS-enforced Windows sandbox for srt

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

sandbox-runtime gains a second Windows backend where the OS enforces filesystem and network policy, and Guardrails fixes a shape rejection that was failing valid text responses.

windows: opt-in MXC BaseContainer backend anthropics/sandbox-runtime

by ig-ant

Setting `windows.backend: 'mxc'` now runs commands through Microsoft's MXC BaseContainer in process via `SandboxManager.spawnSandboxed()`, with no install step and no `wxc-exec` runner. Leaving the key unset keeps `srt-win`, so no existing config changes behavior until you opt in.

fix(server): accept image metadata and refresh OpenAI contracts NVIDIA-NeMo/Guardrails

by Pouyanpi

Text-only upstream responses that carry `image.context_compression` were being rejected by the closed response projections as unsupported shapes. Absent or null `image`, or a closed object with `context_compression: enabled | disabled`, is now accepted across OpenAI Responses and LiteLLM Chat, including streaming snapshots and terminal metadata bags.

fix(agent-rag-governance): reject a negative max_retrievals_per_minute (#4484) microsoft/agent-governance-toolkit

by Shivam Sharma

A negative `max_retrievals_per_minute` used to be accepted by RAGPolicy, then fail every single retrieval with `RateLimitExceededError`, because the limiter treats only 0 as no limit. Policy construction now raises `ValueError` instead, so a bad config fails at startup rather than at query time.

fix(agent-lightning): count only accepted spans in emit_to_store (#4487) microsoft/agent-governance-toolkit

by Shivam Sharma

`FlightRecorderEmitter.emit_to_store` returned the number of spans read from the recorder; it now returns the number the store actually accepted. If you alert or assert on that return value, expect lower counts whenever a span's emit raises or the store exposes no `emit_span`/`add_span`.

docs(hooks): generate hooks server API reference NVIDIA-NeMo/Guardrails

by miyoungc

The hooks server API reference and the experimental proxy reference are now generated from registered FastAPI routes, with contract tests against the running server and the live OpenAPI endpoint left disabled; the macOS arm64 test job moved to macos-15. Documentation coverage only, nothing to migrate.

Quick answers

What shipped in Agent Security on November 6, 2026?
sandbox-runtime gains a second Windows backend where the OS enforces filesystem and network policy, and Guardrails fixes a shape rejection that was failing valid text responses. In total, 7 commits and 7 pull requests landed.
Who contributed to Agent Security on November 6, 2026?
4 developers shipped this update, including Shivam Sharma, miyoungc, Pouyanpi, and ig-ant.
What were the notable Agent Security updates?
windows: opt-in MXC BaseContainer backend, fix(server): accept image metadata and refresh OpenAI contracts, and fix(agent-rag-governance): reject a negative max_retrievals_per_minute (#4484).