Agent security incidents

A dated, sourced list of notable AI agent security incidents and CVEs, with what to do about each.

2026-10-18

593 published advisories for OpenClaw packages

OSV lists 593 advisories across the openclaw and clawdbot npm packages (20 critical, 219 high), 9 without a fixed version yet.

What to do: Stay on a current release and re-check after each update.

Source: OSV.dev

2026-08-31

Stolen API key burns about $600k of credits at METR

METR disclosed that a stolen API key from an agent dashboard was used for roughly $600,000 of credits before anyone noticed.

What to do: Set hard spend limits and per-project keys, and watch usage.

Source: METR

2026-02-19

Microsoft publishes guidance on running OpenClaw

Microsoft's security team advised treating OpenClaw as untrusted code execution with persistent credentials, with identity isolation and runtime controls.

What to do: Run your agent with its own accounts and keys, separate from your main ones.

Source: Microsoft Security Blog

2026-02-12

Dutch data protection authority warns about OpenClaw-style agents

The Autoriteit Persoonsgegevens warned against using agents like OpenClaw on systems with sensitive data.

What to do: Limit what your agent can read, especially email and documents with personal data.

Source: Autoriteit Persoonsgegevens

2026-02-02

CVE-2026-25253: gateway token theft via a malicious link

A crafted link could make OpenClaw/Clawdbot send its gateway token to an attacker (1-click remote control). Fixed in 2026.1.29.

What to do: Update to the latest release and rotate your gateway token.

Source: GitHub advisory GHSA-g8p2-7wf7-98mq

2026-02

17,500+ OpenClaw instances found exposed on the internet

Hunt.io reported more than 17,500 internet-exposed OpenClaw, Clawdbot and Moltbot instances linked to CVE-2026-25253.

What to do: Keep the gateway on loopback or behind token auth, and close port 18789.

Source: Hunt.io

2026-02

Moltbook database exposed agent API keys

Researchers reported an open database at Moltbook exposing about 1.5 million agent API keys and 35,000 email addresses.

What to do: If you used Moltbook, rotate any key you connected.

Source: Wiz Research

2026

ClawHavoc: 341 malicious skills on ClawHub

A poisoning campaign placed 341 malicious skills on ClawHub, built to steal keys and data from OpenClaw users.

What to do: Review skills before installing, pin versions, and give skills their own low-limit keys.

Source: Packet Storm

Check your own setup

Paste your audit and get a plain-English result in minutes. Look-only, point-in-time, and never "green" unless the check finished.

Run the free check