If attackers can remove the guardrails, how do defenders get ahead?
AI can accelerate the work of building cheats, probing systems and adapting to a fix. Game defenders need more than access to consumer models: specialist capabilities, game-specific evaluation and a practical way to put them to work.
An uneven contest.
Open models can be adapted and run without a provider enforcing its safeguards. That gives attackers room to automate parts of cheat development, investigate weaknesses and keep revising their approach as defences change.
Legitimate security teams work under different constraints. A consumer model may restrict the offensive testing a developer needs to carry out on their own systems. Access to general-purpose AI does not necessarily mean access to the capabilities needed for authorised security research.
The concern is the pace of that contest. Model-assisted workflows can keep iterating, while defenders have to investigate each finding, protect legitimate players and ship a reliable response. We think game security needs a stronger answer than playing catch-up with the next cheat.
Give defence a head start.
Verified cyber programmes are expanding access to specialist capabilities. Game studios need a way to turn that access into useful defence: shared priorities, evaluation against game systems and integration with the tools their developers already use.
We are working towards a game-security toolkit that would help teams run authorised investigations, reproduce findings and re-test candidate fixes. We want to develop this with studios and frontier labs, connecting model capability to evidence a developer can act on.
Stronger models are only part of the answer. We need to measure investigation and response time, deployment cost and the reliability of findings. Work on suspicious-play detection would also need evidence about false positives and latency. The aim is to strengthen existing anti-cheat and security work with capabilities that can keep improving.
Our first testbed
We plan to start with a controlled multiplayer inventory and explicit item-transfer rules. A model would inspect code and event traces, test action sequences and produce a reproducible failure. After a developer-approved fix, we would re-run the case and compare with scripted and property-based tests. This is our first way to test the approach.
Further reading
We want to work with game studios and frontier labs on what a stronger defence should look like in practice.
Start a conversation