Anthropic quietly restricts Claude from game reverse-engineering and model training

Anthropic quietly narrows what Claude will do

Anthropic has begun restricting Claude's assistance in at least two sensitive areas — reverse-engineering software and training AI models — without publishing an updated usage policy or explaining the change. According to reporting from Tom's Hardware, relayed by Mashable, Claude now declines or limits requests to decompile, mod or otherwise reverse-engineer video game code. Separately, a machine-learning researcher says Anthropic began blocking his account over work it characterised as "training a competitor," even though he denies ever distilling Claude. Neither change has been officially announced.

Games are the flashpoint

The gaming restrictions arrive after a stretch in which AI tools made game hacking and modding dramatically easier, as Mashable recounts. Leaked Grand Theft Auto VI material spread widely online, prompting Take-Two Interactive to pursue subpoenas against Microsoft and Discord in search of whoever was responsible. AI-assisted decompilation projects followed: fans spliced Nintendo's Mario into Elden Ring and dropped Spider-Man into Batman: Arkham City, while others got World of Warcraft, Call of Duty and Grand Theft Auto: Vice City running in a browser with no say from the rights holders. A PlayStation 5 jailbreak then enabled backward compatibility with older PlayStation titles and cross-platform play, with users running Xbox's Halo 2 and Nintendo's Zelda: Breath of the Wild on jailbroken hardware.

Given the lawsuits and potentially enormous damages such activity invites, Mashable argues it is unsurprising that AI vendors are scrambling to close these pathways. But the timing is awkward. Tom's Hardware reporter Bruno Ferreira points out the irony of frontier AI companies — whose models were built on training data scraped from across the web — suddenly acting as copyright enforcers.

Model training triggers blocks too

The second restriction surfaced in a public complaint. On October 31, Jiahao Shao posted on X that Anthropic had flagged him for "training a competitor." He insists he never distilled Claude: the rollouts in his research came from a strong open-source model, he says, with Claude used only to submit jobs and run analysis. According to Shao, anything touching reinforcement fine-tuning, supervised fine-tuning or the training of larger models now gets him blocked, sees his Opus 5.5 access downgraded to Opus 5, or kills his session outright.

His post drew considerable attention — more than 117,000 views — but it remains a single, unverified account. Anthropic has not commented on it, and the enforcement criteria it describes cannot be independently confirmed. If accurate, it suggests automated systems are policing a fuzzy boundary: helping a researcher orchestrate training jobs can look, from the outside, a lot like building a rival model.

Enforcement by behaviour, not policy

What links the two cases is silence. In both, the policy shift is visible only through the model's behaviour — refusals, downgrades and account blocks — rather than through a documented acceptable-use update that developers could read and plan around. For anyone building workflows on top of Claude, that means the product's practical limits can change without notice.

Mashable also flags a competitive problem. Restricting one assistant does not make the underlying activity disappear; people determined to clone games or jailbreak consoles can simply switch to a model that has not tightened its rules. That risks a race in which the least cautious AI ends up the most popular.

Why it matters

Anthropic appears to be pre-empting legal exposure — from game publishers aggrieved by AI-assisted ports and jailbreaks, and from rivals worried about distillation — by quietly narrowing what Claude will do. For developers, that means the effective rules of the product now shift without documentation, and legitimate work such as research automation or security analysis can trip the same filters as abuse. For the wider industry, it sketches an uncomfortable equilibrium: vendors with the strictest guardrails may simply push their worst-behaved users toward competitors, while paying researchers get caught in the middle.