Skip to content
Agentsstrong signalverified

Gemini CLI 0.58.0 closes a sandbox escape that ran through the Docker socket

Gemini CLI runs untrusted work inside a macOS Seatbelt profile. Until 0.58.0 that profile did not stop a sandboxed process from talking to a container runtime on the same machine, and a container runtime will mount the host filesystem for whoever asks. The fix was merged on August 21, 2026 and reached a stable release on September 1, 2026. It covers the profile builder and all six built-in macOS profiles.

By Redakcija WebAiRadarPublished 2 min readwritten by a model

Source

Gemini CLI v0.58.0

Gemini CLI Releases · Original published September 1, 2026

A sandbox holds only for what the sandboxed process does itself. Gemini CLI confines work in a macOS Seatbelt profile, and until version 0.58.0 that profile left the local container runtime reachable. Asking a privileged container to mount the host filesystem is not something Seatbelt inspects, because the mount is performed by a daemon that was never inside the sandbox.

How the escape worked

On a machine running macOS with Docker Desktop, OrbStack, Colima or Rancher Desktop active, the Seatbelt profile allowed broad process execution, file reading and outbound networking. That combination was enough. A sandboxed process could reach the container daemon socket, start a privileged container, and have it mount the host filesystem through VirtioFS. The files then arrived from outside the Seatbelt boundary, so the boundary did not apply to them.

None of this requires a bug in Seatbelt. Seatbelt restricted the process; it did not restrict what that process could ask another daemon to do on its behalf. That is the general shape of the problem, and it is worth carrying into any sandbox review: the boundary covers direct actions, not delegated ones.

What the profile now denies

The change was merged on August 21, 2026. It applies to the dynamic profile builder in the core package and to all six built-in macOS Seatbelt profiles, and it refuses four categories of access outright.

  • Reads and writes to container daemon sockets, including /var/run/docker.sock, /private/var/run/docker.sock, /Users/Shared/.docker and user-level runtime paths under ~/.docker/, ~/.colima/, ~/.orbstack/ and ~/.rd/.
  • Execution of the container binaries docker, dockerd, docker-compose, podman, colima and orb in the usual install locations, among them /usr/local/bin, /opt/homebrew/bin and application bundles.
  • Mach service lookups for names beginning with com.docker. and for OrbStack's service name.
  • POSIX shared memory whose name carries a Docker-related prefix.

11 days between the fix and the stable release

The commit landed on August 21, 2026, and the first stable release carrying it is 0.58.0, on September 1, 2026. By our own count that is 11 days, during which everyone on the 0.57 line was still running the old profiles. The release notes do not mark the entry as a security fix; it is listed between a symlink handling fix and a write-policy fix, in the same typeface as everything else.

The rest of 0.58.0 is small. Ignore-path handling now evaluates symlinks consistently, the write policy configuration declares its top-level safety checkers, and the agent server clears a stale cancellation error when a new message turn begins.

Sources

BrandsGemini

Related

OPENCLAW 2026.10.232,768characters, and past that a key could slip thr
Agentsmedium signal

OpenClaw 2026.10.2 stops long tool output from slipping credentials past redaction

A bug in OpenClaw's secret masking let a credential through whenever it straddled the boundary between two slices of a long text. The fix ships in 2026.10.2 along with repaired updates from 2026.10.1 and a new compaction default for OpenAI and Anthropic API keys.

OpenClawverified

The OpenClaw web interface with a conversation on the left and the Bits & Bolts reference MCP app open beside it, showing a 3D keycap model and an Attach view + note button.
Agentsmedium signal

OpenClaw 2026.10.1 lets agents search installed skills and changes how sessions message each other

OpenClaw released version 2026.10.1 on October 10, 2026, its first stable release since 2026.9.7. Agents can now search the skills installed on a machine and read only the one a task needs, instead of carrying the whole library in the prompt. Opt-in MCP Apps run beside chat, and full backups can go to an external drive or Cloudflare R2. Three changes can break existing setups: session-to-session messaging, the old required argument, and Workboard data stored before July 2026.

OpenClawverified

CLAUDE CODE2.1.294hooks written as instructions now block what t
Agentsmedium signal

Claude Code 2.1.294 fixes prompt hooks that let through the actions they were written to block

Anthropic released Claude Code 2.1.294 on October 8, 2026. It fixes prompt and agent hooks phrased as instructions, which could allow exactly the actions they were meant to stop. Version 2.1.295 followed the same day with an onFailure setting that makes a crashed or timed-out command or HTTP hook block the action. Releases 2.1.292 and 2.1.293 from earlier in the week closed several permission and sandbox gaps.

Claude Code Releases (GitHub)verified