VACNet Explained: What Valve's AI Anti-Cheat Actually Does

Two CS2 patches in two weeks both touched VACNet, and the reaction to each was roughly the same: VAC finally works.

Neither patch changed VAC. One closed a hole that, in Valve’s own wording, had let malicious players avoid being processed by VACNet. The other added a review tool that a small number of invited players can see, that reviewers say does not ban anyone directly, and that Valve has not announced or documented.

That second one is where most of the confusion lives, so this piece is organized around a distinction worth keeping: what Valve has actually said, what invited players have reported, and what is being assumed by people reasoning about how systems like this usually work. The third category is much larger than the coverage suggests.

VAC and VACNet are two different systems

Valve Anti-Cheat—VAC—is the old one. It runs on your machine and looks for known cheat signatures in memory. It is a pattern matcher. If a cheat is known and unobfuscated, VAC catches it. If it is new, or paid, or rewritten weekly, VAC generally does not.

VACNet is the newer, server-side one, and it does not inspect your computer. It evaluates what your character does—how the crosshair moves, whether accuracy spikes at implausible moments, how shots line up against what you could plausibly see. The version Valve described publicly was built primarily around aiming and shooting behavior. It is a behavioral classifier trained on labeled examples.

That distinction is why a cheat can defeat signature scanning and still get its user banned—and why “undetected” in a cheat seller’s marketing means less than buyers think.

This part is not in dispute, and it is the least speculative thing in this article. Valve engineer John McDonald walked through the approach publicly in a 2018 GDC talk on using deep learning against CS:GO cheating. That talk is still the best primary source on how the system reasons—and a reminder that VACNet is not new.

A cooldown is not a VAC ban

This is the distinction most coverage gets wrong, and it changes what the July patch actually meant for the people it hit.

Valve documents two different consequences, and they are not the same severity:

• Irregular gameplay detection removes you from the match in progress and applies a competitive cooldown of at least one day, locking you out of all game modes for the duration. Valve notes this cooldown also lands on anyone partied with the disconnected player.
• A VAC ban is permanent, applies to the account rather than the match, and Steam Support will not lift it. Valve states that incorrectly issued VAC bans are reversed automatically, which is the only route back.

So the live match cancellations players celebrated after the July patch are the first category—cooldowns—and are documented on Valve’s CS2 cooldown page. Permanent bans are a separate track with separate rules. Saying “VACNet banned him” when the account caught a one-day cooldown is the kind of shorthand that makes the system look either harsher or more toothless than it is, depending on which way you spin it.

The July 29 fix: the clearest thing that happened

Valve’s July 29 release notes carried an unusually direct line—the fix addressed, in Valve’s words, “a bug that allowed malicious players to avoid being processed by VACnet.”

Read that carefully. The affected players were avoiding VACNet processing altogether, so the system was not getting the opportunity to evaluate them.

What Valve confirmed

The patch note itself, and that it shipped on July 29, 2026. That is the entire list.

What is reported but not confirmed

Community analysts and multiple outlets describe the bypass as cheat clients feeding falsified view angles into subtick input history—the data the server uses to reconstruct what happened between ticks. If that record is fabricated, VACNet evaluates a fictional player who happens to aim like a saint.

This explanation is consistent and widely repeated, but it is community analysis, not a Valve description.

Some players also report that ban notifications now name VACNet rather than showing the generic VAC message. That is single-source and unverified.

What players observed

Rage hackers and spinbotters collecting bans within their opening rounds after the patch. Several trackers also reported live match cancellations mid-game.

The speed points toward a backlog—accounts already flagged, waiting on a pipeline that could not reach them—though that is an inference from the timing, not something Valve described.

The Labeling Portal: an early test, not a launch

On August 10–11, 2026—accounts differ on the date—a small patch added the VACNet Labeling Portal, a browser-based review tool at counter-strike.net/vacnet. Invited players see a “V” icon appear in the CS2 main menu linking out to it.

Valve shipped it silently. No blog post, no patch note describing it, no eligibility criteria, no statement of intent. What follows is assembled from screenshots and accounts posted by the small number of players who got in during the first days.

Reported consistently across sources:

• Access is invite-only, and requirements are unknown. Every outlet covering this says so explicitly. Speculation that it is limited to trusted or long-tenured Prime accounts is exactly that—nobody has confirmed a criterion.
• Reviewers watch a clip of a suspect player with X-ray enabled, then label the behavior: aim assistance, wallhacks, movement scripts, or automated accounts such as farming bots.
• Verdicts do not ban the reviewed player. They feed VACNet’s training data. This is the single most consistent claim across every account and the most important thing to understand about the system.
• Valve asks reviewers not to stream, record, or share clips. The reasoning is straightforward—published examples of what VACNet is learning to catch would be a study guide for cheat developers.

Where reports actively conflict: clip length. Some accounts describe roughly three-minute clips; others describe 10–15 second segments of a match. Both are circulating as fact. Neither traces to Valve. It is possible both are true for different case types, and it is possible one is simply wrong.

What is being assumed and should not be stated as fact: that the same clip goes to multiple independent reviewers, and that Valve weights clips by how much reviewers agreed. That would be the standard way to build a labeled dataset—agreement between independent reviewers is what separates reliable training examples from noise—but it is an inference about how such systems normally work, not a description of what Valve built. It has been repeated confidently enough to look like reporting. It is not.

An “Invite Reviewers” control has been described in the portal, which would suggest the pool can grow through nomination. Also single-source, also unconfirmed.

Why training data instead of verdicts

If the reports are right that verdicts do not ban anyone, that is a real departure from CS:GO’s Overwatch, where enough guilty verdicts produced a ban and the community was effectively the jury.

Valve has not explained the change, so the reasoning here is ours rather than theirs, but the tradeoff is easy to see. Community juries scale linearly—a hundred reviewers handle a hundred reviewers’ worth of cases. A model trained on those verdicts applies to every match on every server without further human-review work. Inference still costs compute, but it does not cost volunteer attention, and that is the constraint that actually binds.

The cost is immediacy. Nobody gets banned because you labeled them, and you will never find out what happened to the account you reviewed.

What the community actually thinks

Reception has been split, and the split is informative.

The positive read: this is visible evidence of Valve building enforcement infrastructure after a rough stretch for CS2’s anti-cheat reputation, and doing it without moving to a kernel-level client—which a large part of the playerbase actively does not want. Combined with the July fix landing bans fast, it reads as momentum.

The skeptical read is sharper and worth taking seriously. A recurring community argument is that VACNet has been in development for the better part of a decade and that official servers have functioned as the live testing environment that whole time—with legitimate players absorbing false positives and weak detection while the system iterates. Under that framing, a limited invite portal is not a milestone; it is the same long-running project asking players for more free input.

Other recurring objections:

• Reviewers report being shown blatant cases. If the clips are obvious, human labeling adds little—the hard problem is the ambiguous middle, which is exactly where labels are most expensive and most valuable.
• A limited-access portal is unlikely to change everyday Premier matchmaking in any timeframe players will notice.
• An invite system with undisclosed criteria invites the obvious attack: cheat-adjacent participants deliberately submitting bad labels to poison training. Whether Valve has guards against this is unknown, and it is the strongest technical criticism on the table.

Both camps are reacting to the same underlying fact—Valve shipped something real but told nobody anything about it.

Why this is a slow system by design

The two enforcement tracks move at different speeds, and that is worth separating too. Cooldowns for irregular gameplay can land inside the match itself—that is what players saw after the July fix. Permanent VAC bans have historically arrived in delayed waves rather than instantly, which is a defensible tradeoff: an aggressive automated banner produces false positives, and a mistaken VAC ban is permanent.

For the portal specifically, the gap between “Valve shipped something” and “my games feel cleaner” is likely measured in months, not days. Labeled data has to be collected, a model has to be trained on it, and that model has to ship. Do not expect a July 29 moment—if it works, it works quietly, in a model version nobody outside Valve will see ship.

What a ban wave means for the market

One downstream note if you hold inventory. This section concerns permanent VAC bans rather than the competitive cooldowns described above.

A VAC ban does not delete a player’s skins. It strips their ability to move them. Per Valve’s VAC-ban documentation, items cannot be moved off a VAC-banned account—it can still receive items, but not send them. Every permanent ban in a wave is inventory that quietly leaves circulation rather than flooding it.

The effect on any individual item is almost certainly too small to observe, and anyone telling you a ban wave will move a specific skin’s price is guessing. The direction is the only clear part: enforcement removes supply, it does not add it.

If you do want to see what your own holdings have actually done over that period rather than guessing from headlines, the CS2 Price Checker covers the current market side.

Quick answers

Is VACNet the same as VAC?

No. VAC scans your machine for known cheat signatures. VACNet runs server-side and analyzes in-game behavior. They run alongside each other and catch different things.

Does VACNet ban you permanently?

Not directly, as Valve documents it. Detection of irregular gameplay removes you from the match and applies a competitive cooldown of at least one day across all game modes. A permanent VAC ban is a separate consequence with separate rules.

Can a VAC ban be appealed?

No. VAC bans are permanent and Steam Support will not remove them. Valve says bans issued in error are reversed automatically, which is the only path back.

What did the July 29, 2026 patch change?

Valve’s release notes say it fixed a bug that let malicious players avoid being processed by VACNet. Community analysis attributes the bypass to falsified view angles in subtick input data; Valve did not describe the mechanism. Enforcement against rage cheats became visible within minutes of play afterward.

Does the VACNet Labeling Portal ban cheaters?

Reports from invited players consistently say no—verdicts train future VACNet models rather than affecting the reviewed account. Valve has not confirmed this directly, and has published nothing about the portal at all.

How do I get access to the Labeling Portal?

Unknown. It is invite-only, Valve has published no criteria, and there is no application process. Invited players report a “V” icon appearing in the CS2 main menu.

How long are the review clips?

Reports conflict—some describe roughly three minutes, others 10–15 second segments of a match. Unconfirmed either way, and neither figure traces to Valve.

Do multiple people review the same clip?

Widely assumed, not confirmed. It would be standard practice for building training data, but no source traces the claim to Valve.

Do I lose my skins if I get VAC banned?

They remain in the account’s inventory, but they cannot be moved off the account or sold on the Steam Market.