CVE-2026-43534 is a critical input validation flaw (CVSS 9.1) in OpenClaw before version 2026.4.10. External hook metadata can be queued as trusted system events, so an attacker who supplies a hook name pushes untrusted text into a higher trust agent context. Upgrade to 2026.4.10 or later.
What's affected
The record says OpenClaw versions before 2026.4.10 are affected. The fix is described in a commit and in the project advisory GHSA-7g8c-cfr3-vqqr, which I did not retrieve.
Exposure depends on whether outside parties can send events through hooks to your agent.
How to check if you're exposed
Find the version of OpenClaw you run and compare it with the fixed version above. If you installed it with npm, this works:
npm ls -g openclaw
Otherwise use the version your installer, container tag or deployment record shows. Then list the gateways, channels and webhooks that are exposed to people or systems you do not control, since these issues concern how untrusted input is handled.
What to do now
- Upgrade OpenClaw to 2026.4.10 or later.
- List the hooks and webhooks that your instance accepts, and who can call them.
- Require authentication on every hook endpoint.
- Review agent permissions and remove tools the agent does not need, so a confused agent can do less.
- Check logs for hook names that look unusual.
Why it matters
The weakness is CWE-345, insufficient verification of data authenticity. The system labelled something as a trusted event although part of it came from outside. An AI agent treats trusted system events as instructions it can act on, so mislabelled outside text can steer it.
This is the same trust boundary problem seen in many agent systems. Data and commands must stay separate. The KEV flag is false.
Quick check
Why is input from a hook name a problem?
Because the record says external hook metadata was queued as a trusted system event, so untrusted text gained a higher trust level.
Sources
- GitHub security advisory GHSA-7g8c-cfr3-vqqr (listed in the NVD record, not retrieved)
- VulnCheck advisory (listed in the NVD record, not retrieved)
- Gold page for this CVE
Last reviewed: 19 October 2026.
