Skip to content

AI security

OpenClaw agents kept reading memory after it was disabled

3 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

All about: OpenAI models and products →
Image: openclaw.ai

In OpenClaw, switching off memory access in the configuration did not stop an agent run that had started with it. That run could keep reading memory until it finished. Trail of Bits found the bug in a security audit, and OpenClaw published the results on October 5. If your team runs OpenClaw, the fixes for every actionable finding are on main and ship in the 2026.8.1 and 2026.7.33 LTS stable releases. The audit rated none of its findings Critical and two of them High.

The audit ran through OpenAI’s Patch the Planet initiative, which pairs AI-assisted security research with human review. Trail of Bits used Codex-assisted workflows to look for issues and draft fixes, then checked each finding by hand before filing it. OpenClaw weighed every report against its trust model and release history, fixed and tested the ones it accepted, and coordinated disclosure through private GitHub Security Advisories. The review covered core permissions and the way OpenClaw handles user data across its features.

What the 27 advisories add up to

Trail of Bits filed 27 private repository advisories and 3 standalone hardening pull requests, and all 3 pull requests were merged. Of the advisories, 24 carried a severity rating: 0 Critical, 2 High, 16 Medium and 6 Low. Which two findings drew the High rating has not been specified.

Of those 24, 23 were confirmed vulnerabilities. OpenClaw closed some of them without publishing because they were fixed before they reached a stable release, so fewer than 23 will show up as published advisories. The 24th concerned a bug that was already fixed when Trail of Bits submitted it. The other 3 advisories were defense-in-depth findings, meaning extra hardening. They got no severity rating because they did not cross a documented trust boundary.

Where permissions slipped

The most common problem was permissions getting lost between steps. A request could enter OpenClaw with limited access and then start follow-on work that no longer carried those limits. Sometimes the fix was to carry the original permissions forward. Other times the follow-on work never needed access at all: the post’s example is a filename generator, which has no use for tools.

Aliases caused a second group of bugs. OpenClaw sometimes keeps more than one name for the same thing so older configurations keep working, and a security check could see one name while the system later acted on another. That affected both user identities and feature names. The fix works out exactly which name the system will use before the security policy is applied.

Several reports found a gap between what OpenClaw checked and what it then used. One archive check looked at only part of an archive before the full contents were extracted. In another case, a file path changed after OpenClaw had already approved it. Approval is now tied to the exact file, identity or action that will be used, and any change after that means a new check.

The memory bug belongs to the last group: permissions that change while an agent is working. Some features checked permission when they started and not when they ran later, so turning a feature off did not always reach work that was already running or cached. OpenClaw moved those checks closer to the moment of use, and for work that takes time it may also check again before returning a result. For memory, tools now read the current setting every time they act. OpenClaw says the bug affected only a run that had already been given access.

OpenClaw also drew a lesson for its tests: “Tests need to exercise the real security boundary instead of stopping at the helper where a bug happened to appear.”

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.