A knife can save or kill people AI empowers good people, just as it empowers evil people When the North Korean hacker group Lazarus systematically embeds AI into the attack chain, AI automatically scans for zero day vulnerabilities and generates exploit batches; When WormGPT on the dark web is priced clearly, paying $60 per month can generate phishing emails and malicious scripts in bulk. We have to admit that bad guys have turned AI into low-cost, large-scale weapons of destruction. The only way to face AI attacks armed to the teeth is to defeat magic with magic. 1/A piece of good news The Global Cybersecurity Alliance (GCSA) has just announced that its GCSA Agent achieved a success rate of 91.3% in the CyberGym benchmark test. Compared to the previous CyberGym leaderboard: First place: GPT-5.5-Cyber+OpenAI Agent -85.6% Sixth place: Grok 4.6+Grok Build -79.7% Seventh place: Grok 4.5+Grok Build -79.0% GCSA Agent is based on Grok 4.5 and 4.6. The same underlying model achieved an additional 12 percentage points, surpassing not only Grok's original solution but also OpenAI's specially tuned GPT-5.5-Cyber. This indicates that in security tasks, the design of intelligent agent frameworks and underlying models are equally important, or even more important. What did GCSA Agent do correctly? GCSA did not simply throw the vulnerability description to the large model, but instead built an autonomous closed-loop vulnerability mining process: Retrieve relevant code from thousands of files and millions of lines of code Identify attack surface Propose vulnerability hypothesis Generate test inputs (PoC candidates) Execute in a real environment Analyze crashes or execute feedback Iterative optimization of PoC based on feedback Verification: Can be triggered before patching, does not trigger after patching Its breakthrough lies not in the model (still using Grok), but in the framework: how to retrieve code, propose hypotheses, parse crash logs, control iteration pace, and when to decisively abandon the current approach. The model is the engine, and the framework is the driving system. GCSA did not build a stronger engine, but created a better driving system. 3/And then what? Grades are just the starting point, and real-world safety work goes far beyond that. What GCSA really wants to do is involve AI agents in the entire security lifecycle: Discovering vulnerabilities independently without anyone telling it where they are Not only generating PoC, but also analyzing the scope of influence Not only can it report problems, but it can also suggest and even generate repair solutions Not just single point combat, but also continuous monitoring of new risks brought by code changes When attackers have already used AI to reduce the cost of committing crimes to near zero, the only feasible way out for the defense side is to let AI patrol for you 24 hours a day without interruption.