Is 360's AI Security Agent Worth Having a Security Team Use?
Community Discussion · General

Is 360's AI Security Agent Worth Having a Security Team Use?

Old Ye from BCGOld Ye from BCGOct 52026/10/05 268 views

I spent two days trying out the AI security agent in 360 Security Cloud. First, my identity: I do strategy consulting, not a security engineer. What I usually show clients is ROI and implementation paths. This time I opened a trial account myself and fed in a set of simulated alert data to run.

What does it do. Security operations work is like this: every day hundreds or thousands of alerts pop up, analysts look at them one by one, judge which are real and which are false positives, then decide whether to handle them. An AI security agent brings in a large model to let it analyze first for you, giving judgments and handling suggestions. 360 Enterprise Security Cloud itself is a SaaS foundation released back in 2022, and this time the agent layer was added on top.

The console is quite clean. On the left are modules for assets, alerts, and agents; in the middle is the alert list; click open an alert and on the right you can see the agent's analysis process.

It won't just throw you a line like "this is malicious behavior" and call it done.

360 itself mentioned one point, and I verified it this time: for security operations agents, all analysis results ultimately have to be handed to experts for supervision and review, so it must output a complete evidence chain — which data was called, why this judgment was formed, which evidence supported or negated the hypothesis, and what the confidence basis for the conclusion is. I clicked open a few and indeed could see which data sources it called and which historical records it compared.

For comparison, I used the same thirty-some alerts, running them once with a general large model pasting logs directly, and once with this agent.

Dimension General large model pasting logs 360 AI security agent
Initial setup Almost zero Need to connect data sources, configure permissions; took me half a day
Analysis output Conclusion-focused, reasoning sometimes there sometimes not Conclusion plus call records plus evidence chain
Stability Same batch of data gave different results twice Official claim: same Skill run about 100 times, over 95% stably follow instructions
Handling actions Can only suggest Can connect to tickets and handling processes
Suitable scenarios Scattered troubleshooting, temporary analysis Daily operations with fixed processes

Where did it get stuck. It got stuck in the early stage. Almost the same feeling as setting up monitoring on the SAP side — 70% of the work is clarifying permissions and environment. This product needs to read your logs, assets, and endpoint data. If permissions aren't sufficient, its analysis is just guessing; if too much is given, compliance won't pass. I basically spent the first day doing this, and actually running alerts was a second-day afternoon thing.

The pleasant surprise is the readability of the evidence chain. I thought it would be a pile of jargon, but actually it breaks down why it judged this way into several steps, and even someone with my non-security background can follow. For a consultant who needs to report to management, this value is not small.

Let me also say what I'm dissatisfied with. The access threshold is not low; for small teams without dedicated security staff, just the configuration can scare them off. Its value strongly depends on your existing data quality; if the logs themselves are incomplete, what it gives can only be a reference.

I saw a review saying that 360's security lobster set cut agent deployment time from 6 hours to 10 minutes. This time Security Cloud's approach is consistent — pushing the threshold down. But for security operations, after installing it there's still the governance hurdle to pass.

Conclusion: It depends. Suitable for medium-to-large enterprises that already have a security team, ticketing and logging systems, and are willing to spend people on permission cleanup, using the agent as an analyst's assistant. Not suitable for small teams of three to five people expecting out-of-the-box use, and not suitable to treat it as a decision-maker — what it gives is still hypotheses, and humans have to review.

Looking ahead, the focus of competition for security agents will land on evidence chains and audit trails, and model capability will actually not be the key. Google Cloud is also pushing the same direction, letting agents classify, investigate, and respond at machine speed, but not handing over control. 360 is on this path; what remains is to see the implementation.

2 replies

?
Ctrl + Enter to reply
Long Ji
Long JiOct 5

I actually tried it out. The point that 70% of the workload is permissions sorting is so true. Small teams simply can't handle it.

Sleepy
SleepyOct 6
Reply to Long Ji

I got stuck on the permissions thing for half a day too, but OP says small teams can't handle it — I actually think incomplete logs are the real dealbreaker. Feed them in and it can only guess blindly.