Ask ten CS2 players what VACnet is and you will get ten answers, most of them partly wrong. It is not a client you install. It does not scan your hard drive. It is not the thing that ends your match when someone gets caught mid-game, though it is closely related. VACnet is a machine-learning model that runs on Valve’s servers and judges behaviour rather than software, and to understand why Counter-Strike 2’s anti-cheat works the way it does, you have to understand where that model got its training data.
GDC 2018: the announcement
VACnet was introduced publicly by Valve engineer John McDonald at the Game Developers Conference in 2018. The core pitch was a rejection of the classic approach. Signature-based VAC asks a question about your computer: is there a known cheat process hooked into the game? VACnet asks a different question, about the game itself: does what this player is doing look like what a human does?
Because that analysis happens server-side, on the streamed gameplay data, it does not need to inspect your machine at all. That single design decision solves several problems at once. There is less room for false positives caused by whatever software a player happens to be running. There is nothing for a cheat to hook into and hide behind. And the detection can be updated on Valve’s side without pushing anything to a million clients.
The Overwatch training set
The most interesting part of the original design was the labels. VACnet was trained on Overwatch verdicts — thousands of cases where trusted community investigators had watched a suspect’s demo and decided whether the aim looked assisted or the vision looked unnaturally informed. The system Valve built was, in effect, a learned compression of what experienced human reviewers had already decided.
That lineage explains a lot about CS2’s anti-cheat arguments. Valve spent years improving an automated system using human judgement as its ground truth, and when CS2 launched without the public Overwatch queue, the community felt the loss precisely because the human feedback loop was invisible. The machine kept running, but nobody could see what it had been taught.
It also explains the recurring community complaint that VACnet seems cautious. A model trained to agree with conservative human reviewers inherits their conservatism. Where an investigator would shrug and vote “not enough evidence,” the model learned to withhold conviction, which is exactly why a subtle cheater can survive in Premier for weeks while an obvious one gets caught quickly.
Integration with Overwatch and the Trusted Mode era
By mid-2018 Valve was describing a pipeline rather than a product: VACnet pre-screened suspicious gameplay and fed the interesting cases into Overwatch review, raising both the volume and the quality of the cases humans saw. That arrangement ran through CS:GO’s Trusted Mode era, when Trusted Mode was made mandatory in June 2021 and third-party injection into the game process was restricted outright.
CS2 inherited the pipeline and rebuilt the enforcement layer around Source 2 and sub-tick. VAC Live became the real-time arm, able to end a match on the spot, while VACnet remained the analytical background process deciding which accounts deserve scrutiny and, ultimately, a delayed wave ban.
The sub-tick leak, and what it proved
The clearest demonstration of VACnet’s dependence on data quality came in 2026. Valve’s update of 28–29 July 2026 included a deceptively small fix: a bug that had allowed malicious players to avoid being processed by VACnet entirely. The reported mechanism was spoofed sub-tick input, with rage cheats sending fabricated view angles into the input history so the server saw plausible numbers while the rendered gameplay did something far less plausible.
That incident matters for two reasons. First, it explains why some of the most obvious cheaters in the game went unbanned for months: they were not being misjudged by the model, they were never being fed to it. Second, it shows that VACnet’s Achilles heel is not its intelligence but its inputs — a server-side model is only as good as the data the client is allowed to report.
VACnet 3.0 and the labeling portal
Community shorthand through late 2026 has drifted toward calling the current generation “VACnet 3.0,” a loose label covering the sub-tick fix, the tightened live detection and, most significantly, the return of human labels. In August 2026 Valve opened an invite-only portal where selected reviewers watch clips and classify behaviour into aim assistance, wall hacks, automated bunny hopping and bot-like farming — explicitly to generate training data for future VACnet models.
That closes the circle beautifully. The system Valve trained on community verdicts in 2018 is again asking the community for verdicts, this time not as a jury but as a labelling service. Whether the next model is better than the last depends almost entirely on how good those labels turn out to be.

Leave a Reply