[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 (IBM security bulletin, Nov 12, 2026; CVE record); no exploitation reported

IBM ContextForge MCP Gateway CVE-2026-110279: any logged-in user could swap the gateway's stored upstream credential for their own via one header, bypassing per-user permissions and audit attribution

  • Fix: upgrade ContextForge MCP Gateway to v1.1.1.
  • Until then: drop the X-Upstream-Authorization header at your reverse proxy for every MCP endpoint, and keep the gateway off untrusted networks.
  • Look back: in proxy logs, search for client requests carrying X-Upstream-Authorization; legitimate clients of a gateway that holds the upstream credential should not need to send it.
  • Why it matters: an MCP gateway's job is to hold credentials so agents and users don't; any client-controlled header that reaches upstream auth undoes that.

Sources: IBM security bulletin: ContextForge MCP Gateway CVE-2026-110279 (Nov 12, 2026) · GitHub Advisory GHSA-fp2v-98c9-h9m3 (CVE-2026-110279) · IBM/mcp-context-forge fix PR #7341