Claude and Codex Users Report Blocks on AI Reverse Engineering of Their Own Devices

The practice, often called Reverse Engineer Anything, has spread in recent weeks as hobbyists use coding agents to decompile firmware. They then replace manufacturers' software, revive internet-connected devices that companies have stopped supporting, or strip advertising from televisions. 404 Media said several users reported triggering safety checks in Claude Code that stopped decompilation. A thread on the project's Discord server last week said Claude was "refusing to decompile anything."

On the OpenAI side, a user who goes by log4j on X said their account was closed for violating rules on "cyber abuse" after using Codex to reverse engineer Samsung televisions. Other users told 404 Media they had been warned or had their accounts 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, has also shut down, the outlet reported.

OpenAI engineer Tarun Gogineni said on X that the restrictions were not intended. "These blocks are mostly accidental and a result of difficult decision boundaries; there is no prohibition against reverse engineering with codex. it's getting fixed," he wrote. Anthropic has not publicly explained the Claude refusals.

For companies that rely on frontier models in security work, the episode shows a practical risk: guardrails meant to stop malicious hacking can also catch legitimate research, and they can change without notice. Reverse engineering is routine in penetration testing, malware analysis and interoperability work.

Matt Brown, principal consultant at Brown Fine Security, told 404 Media that the blocks he had hit lined up almost exactly with the labs' "cyber" category filters. He said he believed the labs' cyber guardrails were "erring on the side of being too strict" under pressure to keep their models out of criminal hacking. He said he is now buying his own hardware to run models locally. "I really don't want to invest tooling and energy into an ecosystem that will eventually lock me out," he said, adding that open-weight models, mostly from China, are improving.

OpenAI has not said when its fix will be in place.