Frontier-model guardrails are pushing legitimate device hacking back to local AI

AI-assisted reverse-engineering briefly made ownership more practical. People used coding agents with tools such as Ghidra and IDA to inspect firmware, revive abandoned internet-of-things devices, replace smart-TV software and document undocumented hardware commands. The technique did not eliminate decompilers or expertise; it let an agent coordinate them, explain evidence and produce a clean-room implementation at dramatically lower cost.

Users of Anthropic’s Claude and OpenAI’s Codex now report that similar work triggers cyber-safety refusals, warnings or account restrictions. The Reverse Engineer Anything community has documented Claude refusing decompilation requests, while a researcher who used Codex on Samsung televisions said OpenAI banned the account for ‘cyber abuse’. Decomp.wtf, a service intended to make software and game analysis easier, has shut down. These are user reports rather than a systematic audit, so they establish inconsistent enforcement—not a universal prohibition.

OpenAI engineer Tarun Gogineni said reverse-engineering is not banned in Codex and that many blocks were accidental consequences of difficult safety boundaries being fixed. That distinction matters, but it does not remove the operational problem. A security researcher building tools around a hosted model cannot know whether the same lawful workflow will run tomorrow or lock the account. Policy uncertainty becomes a form of platform risk.

The safety boundary is genuinely difficult. The same capabilities used to recover firmware, audit a product or achieve interoperability can expose credentials, defeat access controls or clone proprietary software. A recent legal paper argues that existing copyright, trade-secret, anti-circumvention, computer-misuse and contract doctrines assumed reverse-engineering would remain expensive and specialized. Agents remove that practical filter without resolving which uses are lawful.

The likely result is not the end of the technique but a change in where it happens. Researchers are moving toward open-weight models such as GLM that run locally and cannot revoke access mid-investigation. That makes the work more private and predictable, but less accessible to ordinary device owners and harder for model providers to monitor. Frontier labs can reduce misuse by enforcing cyber boundaries; when those boundaries cannot distinguish ownership repair from intrusion, they also create the demand for uncensored local systems that sit outside their controls.