VACNet 2: Valve Is Training CS2 Anti-Cheat With Human Help

Wed, 12 Aug 2026 00:00:00 +0000

Anti-cheat is one of the most painful parts of competitive games, and Counter-Strike 2, also known as CS2, is a good example of that.

In the last few weeks, I have been playing CS2 again and I noticed a lot of suspicious players. Some matches had weird timing, strange crosshair placement before seeing an enemy, and players who seemed to know too much. You know that feeling when the game stops being about winning the match and becomes about asking: “is this guy cheating?”

But recently, at least from my experience, this started to get a little better. It is far from perfect, but I have seen fewer obvious cases than before.

At the same time, Valve, the company behind Counter-Strike, launched the . It is an invite-only portal where selected players can review short gameplay clips and classify suspicious behavior. According to , these reviews are not used to instantly ban players. They are used as training data for VACNet, Valve’s machine-learning anti-cheat system.

And this is the interesting part for me.

This is not exactly the old Overwatch from CS:GO coming back. Overwatch was a community review system where trusted players watched recorded matches and their verdicts could help punish cheaters. In this new version, the goal seems to be different: humans are helping Valve teach the anti-cheat what cheating looks like in CS2.

That makes sense because CS2 is not CS:GO. The game engine changed, movement changed, the way the server processes player actions changed, and many small details of the game changed too. If VACNet was trained using CS:GO data, that data is not perfect for CS2 anymore. The system needs new examples from the new game.

As a trusted member, I was invited to participate in this portal, and the process is simple but important. You watch a short clip, try to understand what is happening, and label the behavior if there is something suspicious. It can be aim assist, where software helps the player aim; wallhack, where the player can see enemies through walls; auto bunnyhop, a movement cheat that helps the player keep speed while jumping; farming bot, an automated account used to farm rewards or stats; or sometimes you just skip because the clip is not clear enough.

For me, this is the right way to use humans in an anti-cheat system. We are not there to be angry and ban people manually. We are there to help create better data.

Good data is what makes the model better.

If many trusted reviewers label clips carefully, VACNet can learn patterns from CS2 itself. It can learn how a normal player moves, how a cheater behaves, and where the difference is. Later, the system can use that knowledge at a much bigger scale than humans ever could.

Of course, this also has problems.

If the labels are bad, the model learns bad things. If the clips are too easy, the system may only get better at catching obvious cheaters and miss more subtle ones. And if players expect instant bans, this approach can feel slow.

But I still think the idea is good.

Anti-cheat is not only about detecting a cheat program. It is also about behavior. How someone aims, reacts, moves, checks dangerous positions, and repeats decisions over many matches can tell a story. Machine learning can be useful here, but only if the training data is good.

That is why the VACNet portal is important. It shows Valve is not only trying to ban cheaters today, but also trying to build a better system for the future.

I do not think VACNet 2 will magically fix cheating in CS2. No anti-cheat can do that.

But after seeing so many suspicious matches, and now helping label clips myself, I feel a bit more optimistic. Maybe the future of CS2 anti-cheat is not only AI or only humans.

Maybe it is both: humans helping train the system, and the system helping protect the game at scale.

Machine Learning |

VACNet 2: Valve Is Training CS2 Anti-Cheat With Human Help

References