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

← Context & Harness Attacks · markdown

official (GitHub advisory GHSA-j7mp-839j-jrm7; CVE-2026-126196 published to NVD Nov 12, 2026); fixed in 1.1.0; no exploitation reported

pgEdge Postgres MCP Server: 'read_only' could be bypassed with ordinary transaction SQL, letting an agent create, alter and drop tables (CVE-2026-126196, CVSS 8.1; fixed in 1.1.0)

  • Upgrade: pgEdge Postgres MCP Server to 1.1.0 or later (released Sept 4, 2026; the CVE was only published Nov 12).
  • Real fix for any database MCP: connect agents with a role that can't write (for example one granted only pg_read_all_data or SELECT on specific schemas). A tool's read-only flag is a convenience, not a security boundary.
  • Prompt-injection angle: an agent that reads untrusted text and holds a 'read-only' DB tool could be steered into destructive SQL on unpatched versions.

Sources: pgEdge: GHSA-j7mp-839j-jrm7, read-only mode bypass by SQL statement smuggling · NVD: CVE-2026-126196 · pgEdge Postgres MCP Server v1.1.0 release