Skip to main content
NOV 2026 New identity Latest guide updates. Latest: AI Design Patterns Local vs cloud AI Grill me Changelog →

Security hub

Secure Claude Code and AI Agents

Start with a local repository check, review current evidence, then follow the security path that matches your task.

Threat DB v2.34.0Updated November 5, 2026

Local check

Start with your repository

Install AgentSec and run its current alpha scanner locally. This website does not receive or inspect your files.

AgentSec 0.1.0a0
agentsec scan /path/to/repository --format json --redact

01 / Evidence

Current evidence, with its date

These values come from the bundled AgentSec feed and describe that database snapshot. They are not live measurements of the entire ecosystem.

Skills Scanned
3,984
Have Flaws
36.82%
Critical-Risk
534
Malicious Payloads
76
CVEs & Advisories
243
Exposed Servers
~1,000

02 / Local check

Verify your repository with AgentSec

The scanner checks bounded repository inputs against bundled intelligence. A clean result is not proof of safety.

3. Respect the boundary

Current detectors do not cover every host-level persistence mechanism, Git-history witness, or runtime behavior.

3 detector records
  • clawhavoc-skill v1: Detect repository-local agent-skill evidence associated with the documented ClawHavoc fake-prerequisite campaign.
  • fakegit-archive-lure v1: Detect repository-local indicators of the FakeGit / SmartLoader campaign that poses as agent tools: sourced artifact hashes, README links to a committed installer archive, and archives inside agent configuration directories.
  • shai-hulud-keyv v1: Detect repository-local indicators associated with the August 2026 Shai-Hulud Keyv/cacheable npm campaign.

04 / Latest

Latest sourced intelligence

Three recent feed events, ordered by their canonical date. Follow the source before making a time-sensitive decision.

Researcher lists 73 GitHub repositories with committed lure ZIPs

Thomas Roccia published a gist of 73 GitHub repositories posing as Claude skills, Codex tools, MCP servers and OpenClaw projects. Each commits a ZIP that its README links; 51 were live at a 2026-11-04 re-check. Roccia reports that SKILL.md and settings files are clean and that two repositories store the ZIP under .claude. Per-row confidence differs and is preserved in the threat database.

Claude Code Action runs MCP servers from a pull request .mcp.json

The action checked out attacker-controlled PR head branches, read .mcp.json from the working directory and enabled all project MCP servers. A pull request carrying a malicious .mcp.json could run code on the Actions runner and expose workflow secrets when a privileged user triggered the action on that PR.

GitHub MCP Server unauthenticated crash on completion/complete

A completion/complete request without a ref field dereferenced nil and crashed the server before authentication, letting any client cause a denial of service.

Open all threat intelligence β†’

05 / Checklist

Three checks before you continue

Use this quick pass now. The full hardening route keeps the complete workstation and team checklist.

0 of 3 complete

Repository
Runtime
Team controls

Open the complete checklist β†’