[ctx]ctxseccontext security: attacks on the harness and context layer of AI agents, plus exploited-now news, hunting and playbooks.

← Context & Harness Attacks · markdown

advisory (GitHub Advisory Database GHSA-2gc9-c85v-5vwr, critical, published Oct 26); no exploitation reported

Google's MCP Toolbox for Databases 1.2.0-1.4.0: CVE-2026-93328 lets an unauthenticated caller copy host files to an outside Cloud Storage bucket and overwrite local files

A critical advisory published Oct 26 covers the cloud-storage source in Google's MCP Toolbox for Databases (googleapis/mcp-toolbox). Path traversal plus missing authorization lets an unauthenticated remote attacker use crafted tool-call parameters to exfiltrate host files to an external GCS bucket and overwrite arbitrary local files. Affected versions are 1.2.0 through 1.4.0.

The current release is 1.15.0 (Oct 23), so most up-to-date installs are clear. The risk is pinned or forgotten Toolbox servers, often deployed as a sidecar for an agent. Check: find Toolbox binaries and images in that range, especially any with a cloud-storage source configured or reachable from outside, and upgrade. Treat a Toolbox endpoint like any other unauthenticated API: keep it off the internet and put auth in front of it.

Sources: GitHub Advisory Database: GHSA-2gc9-c85v-5vwr (CVE-2026-93328) · CVE record: CVE-2026-93328 · googleapis/mcp-toolbox releases