What happened
In controlled testing, a cross-origin WebSocket client sent a desktopCommand frame that added an attacker-chosen stdio MCP server entry to the victim's Cline settings and received an ok response.
2026-09-24
HighPrimaryCVE-2026-59723 let a malicious website connect to the local Cline Hub dashboard, alter MCP settings, and steer auto-approved agent sessions.
What happened
In controlled testing, a cross-origin WebSocket client sent a desktopCommand frame that added an attacker-chosen stdio MCP server entry to the victim's Cline settings and received an ok response.
Why it matters
The demonstrated impact was persistent unauthorized MCP configuration modification. The advisory documents potential local file, command, network, and credential access when the injected server is activated or a configured model drives an auto-approved session.
Missing authorization check
Independent authorization for MCP server changes and consequential tool execution, bound to the caller, command, target, and session.
Would PP block it?
Protected merge, deployment, publication, credential use, and infrastructure operations could require a receipt scoped to the exact action and target. A hijacked Cline session or injected MCP process would not possess that independent authority.
Incident analysis
2026-09-24
GitHub publishes GHSA-3cj3-hqcr-g934 and CVE-2026-59723 with dynamic reproduction details.
2026-09-24
Cline 3.0.30 is identified as the first patched release.
Authorization boundary
This incident is categorized as Tool execution / MCP. The relevant Permission Protocol gate is Tool-Call Gate. The read is conditional: the block only applies where the real action boundary is routed through a gate.
Permission Protocol cannot repair Cline's missing Origin validation or prevent local settings writes, but an external Tool-Call Gate can deny consequential downstream actions that lack a separately authenticated receipt.
Related incidents and controls
Agentjacking: Attackers inject malicious instructions into fake Sentry errors via public DSN, hijacking Claude Code, Cursor, and Codex with 85% success rate across 2,388 organizations
CVE-2026-82533 Let a DeepSeek Harness Agent Disable Its Own File Sandbox Without Human Approval
Start small
This incident maps to Tool-Call Gate. Start with the boundary that controls the actual action, then require a signed receipt before execution.