360龙虾卫士

What 360 AgentSafe is

360 AgentSafe, branded in Chinese as 360 Lobster Guard, is an endpoint security product for AI agents that guards the local machine where a desktop agent such as OpenClaw runs. It is the defensive counterpart to the agent wave, and its security framing is unusual for this category in that it quotes government risk advisories.

What you can do with it

Two engines work together: a local active-defence engine backed by the vendor's security brain scans and verifies skills in real time, and a cloud identification engine uses a model to judge logic anomalies and poisoned skills, so the loop is scan, judge, block. Protection spans files, network, processes and screen across a set of capability areas said to cover many attack scenarios, including blocking reads of private keys, stopping destructive delete commands, marking malicious skills before install, blocking prompt injection from web pages, scanning for known agent vulnerabilities, gating document-directory access, and pausing screen access when payment pages appear. It advertises coverage for a large number of known agent vulnerabilities, a trust model that grants sandbox rights by signature provenance, and a permission console for file, system-execution and skill-calling scopes, with skill sandboxing that keeps extensions away from the core system.

Who it is for

It suits individuals and IT teams running desktop agents that can touch their files, terminal and screen and who want a safety net around them.

A few things worth knowing

No price, trial length or licensing model is published, so request a quote and confirm whether a free tier exists, and because it defends the machine it is worth installing deliberately rather than casually. Two honest caveats come from the page itself: the process-control and outbound-traffic features are not live yet, and despite quoting advisories from the national computer emergency response team and a government vulnerability platform, the vendor lists no compliance certification, privacy policy link or data-handling statement. Ask where telemetry goes before enabling cloud judgement, and treat it as a hardening aid rather than a guarantee.