Skip to main content

Claude and Codex Refuse Reverse Engineering: Why Users Go Local

Users report that Claude Code now often refuses to decompile software and that some Codex accounts were restricted or banned after reverse engineering hardware. An OpenAI engineer says the Codex blocks were mostly accidental and are being fixed. Anthropic has confirmed only a narrower rule against recreating existing games. Some researchers are moving the work to local open-weight models instead.

On this page

Reports of refusals and bans pile up

AI-assisted reverse engineering became one of the more popular uses of coding agents this year. People used Claude Code and Codex to install new operating systems on smart TVs, turn baby monitors into assistants, and revive abandoned smart home devices. That work is now running into guardrails.

404 Media reports that several Claude Code users triggered safety checks when they tried to reverse engineer software. A thread in the Discord for Reverse Engineer Anything, a popular tool for this kind of work, said Claude was refusing to decompile anything, and a highly upvoted Reddit post warned that the tool was unlikely to work with Claude. On the OpenAI side, one user who used Codex to reverse engineer Samsung TVs said their account was banned for violating rules on "cyber abuse", and others said they were warned or restricted after asking Codex to analyse hardware or their own code. Decomp.wtf, a project that made it easier to reverse engineer software and games, also shut down in recent days.

What the companies have actually said

Neither company has announced a formal policy against reverse engineering. OpenAI engineer Tarun Gogineni posted that the Codex blocks were "mostly accidental", the product of difficult decision boundaries, and that there is no prohibition on reverse engineering with Codex. He said it is being fixed.

Anthropic has confirmed something narrower. Claude Code's instructions now tell Claude not to build a recreation of a specific existing game, even if it is renamed or described as original, while still allowing original games, genre games, and games built from a user's own designs. Addy Osmani, who works at Anthropic, said the aim is to keep Claude from being used to infringe on someone else's game, which the usage policy already forbids, and asked people to report refusals of work that is their own.

Reports of broader refusals go beyond that rule. Shacknews notes users claiming that non-malicious, generic reverse engineering tasks were also blocked, and PCWorld reports that Claude has reportedly started refusing prompts to strip bloatware from smart TVs. Anthropic has not said whether those refusals are tied to the game rule or to separate cyber safeguards.

Security researchers see a cyber filter at work

Matt Brown, a principal consultant at Brown Fine Security, told 404 Media that his reverse engineering blocks were "more or less 100% correlated" with refusals in the cyber category. His reading is that labs are being overly strict because of pressure to stop criminal hacking, and the same skills that unbrick a gadget can also find a vulnerability.

That tension is real. Decompiling firmware is how researchers find hidden commands and security holes, and it is also how attackers find them. A classifier tuned to catch the second group will inevitably catch some of the first.

Why this pushes the work onto local models

Brown found hidden Bluetooth commands in an internet-connected gas monitor on his propane tank using GLM 5.3, an open-weight model. He argues that local and open-weight models matter precisely because frontier guardrails can change without warning on the same model, and he does not want to build tooling in an ecosystem that could lock him out.

That is the practical lesson for anyone whose work depends on an AI tool behaving consistently. A hosted model's behaviour is set by its provider and can shift overnight through a system prompt or classifier update. A model running on your own hardware keeps the same weights until you change them. The trade-off is capability and cost: the strongest open-weight models need serious hardware, and smaller local models are less reliable on hard decompilation tasks. For people reverse engineering devices they own, though, that trade-off is starting to look worth making.

CuriousLM runs supported AI models locally on your device. Try CuriousLM.