RosmarOps

RosmarOps Observatory

Agent security, day by day

What changed in the security of the AI coding agents we follow since the last check, from their official sources and public registries. Up to five changes a day, one line each, with its source. Days without changes are listed too.

  1. Agent security today: 17 November 2026

    Claude Code 2.1.326, dated 16 November 2026, warns at startup about incomplete --allowedTools, --disallowedTools or --tools entries and explains when managed-mcp.json stops MCP servers loading, according to its changelog.

  2. Agent security today: 16 November 2026

    Qwen Code v0.27.4, dated 15 November 2026, lets Managed Agent sessions use granted local directories and a Managed Web Terminal without storing credentials or paths, according to its changelog.

  3. Agent security today: 15 November 2026

    No changes found in the 43 sources checked today. Not checked: Comet.

  4. Agent security today: 14 November 2026

    GitHub Copilot CLI 1.0.111, dated 13 November 2026, lets enterprise-managed permission rules for Shell, Bash and PowerShell use regular expressions, according to its changelog.

  5. Agent security today: 13 November 2026

    OSV published MAL-2026-19204 on 13 November 2026, listing the PyPI package my-mcp-server-driving-license-ocr as containing malicious code.

  6. Agent security today: 12 November 2026

    OSV published MAL-2026-19148 on 12 November 2026, listing the npm package @rfewqa-fes/mnm-doc-mcp as containing malicious code.

  7. Agent security today: 11 November 2026

    CVE-2026-83953, published in NVD on 10 November 2026, affects GitHub Copilot: a command injection flaw lets an authorized attacker disclose information locally.

  8. Agent security today: 10 November 2026

    GHSA-5xjx-9qx6-jcrv, published in OSV on 9 November 2026, affects the npm package @anthropic-ai/claude-code: a race condition in the Edit tool can write files outside the workspace.

  9. Agent security today: 9 November 2026

    CVE-2026-18484, published in NVD on 8 November 2026, affects IBM ContextForge MCP Gateway 1.0.0 through 1.0.11: tool name collisions from slug normalization can cause a denial of service.

  10. Agent security today: 8 November 2026

    CVE-2026-123493, published in NVD on 8 November 2026, affects Vikunja 2.4.0 before 2.7.1: disabled or locked users can keep access, including over MCP, through their bot users' API tokens.

  11. Agent security today: 7 November 2026

    CVE-2026-102610, published in NVD on 6 November 2026, affects fast-agent before 0.7.21: fast-agent serve exposes its MCP endpoint on all interfaces without authentication.

  12. Agent security today: 6 November 2026

    CVE-2026-110930, published in NVD on 5 November 2026, affects Codex Desktop for Windows: quotation marks in Explorer project paths can add startup arguments that run code before sign-in.

  13. Agent security today: 5 November 2026

    The review could not check any source today.

  14. Agent security today: 4 November 2026

    CVE-2026-76894, published in NVD on 3 November 2026, affects CKAN MCP Server before 0.4.112: its DataStore table interface accepts window messages without checking their origin.

  15. Agent security today: 3 November 2026

    CVE-2026-116635, published in NVD on 2 November 2026, affects AWS Serverless MCP Server before 0.2.1: path traversal in its deployment status component reaches JSON files outside its directory.

  16. Agent security today: 2 November 2026

    CVE-2026-120323, published in NVD on 2 November 2026, affects MCPHub, an MCP server hub, up to 1.0.45 and 1.1.7: its passthroughHeaders parameter allows information disclosure; 1.1.8 fixes it.

  17. Agent security today: 1 November 2026

    No security changes found in the sources we checked.

  18. Agent security today: 31 October 2026

    CVE-2026-62863, published in NVD on 30 October 2026, affects MoLing, an MCP server, before version 0.5.0: in SSE mode its endpoints accept requests without authentication.

  19. Agent security today: 30 October 2026

    No changes found in the 14 sources checked today. Not checked: OpenAI Codex.

From the vendors' official sources and public registries, each line with its date and source. Not a security guarantee or advice.