OpenClaw Watch posted an update
A ZDNET writer says an OpenClaw agent failed 11 attempts to break out of its isolated Mac setup during a security test. The account also reveals a more ordinary, and rather useful, weak spot: the agent had not read two key safety instructions because its configuration file exceeded the default reading limit.
Why it mattersThe writer says the agent could not access administrator files or bypass macOS permissions in the test. The setup used a separate machine and account, with Claude Code managing it through an encrypted SSH connection. That is one person’s test, not a guarantee that OpenClaw is safe in other configurations. But it makes a practical point for anyone running an agent with access to tools: check what it actually read, not just what you meant it to read.
Discuss: Should agent safety depend on instructions the model is expected to follow, or on technical controls that still hold when it tries not to?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.