Claude Code and Codex Clamp Down on Reverse Engineering: OpenAI Banned Accounts by Mistake

You bought a smart TV, and it shows ads right in the menu. You used to be able to point an AI agent at the firmware, ask it to figure out what is running there, and strip out the junk. Now the agent may answer "I can't." And with Codex, you can even lose your account. This is reported by 404 Media.

What happened

The trigger was the popularity of the Reverse Engineer Anything project. It uses AI agents to analyze programs without source code. Enthusiasts have applied such tools to a range of tasks: replacing the stock shells of smart TVs, removing ads, reviving devices abandoned by their manufacturers, and dissecting closed protocols.

After the surge of interest, Claude Code users began reporting that safeguards kick in when they attempt decompilation. In the project's community, people complain that Claude refuses tasks it used to perform without question.

The situation with Codex is harsher. An X user going by log4j said OpenAI blocked his account for "cyber abuse" after he reverse-engineered Samsung TVs. Others received warnings and access restrictions while analyzing hardware and even their own code. It later emerged that log4j's account had been unblocked as early as November 3, and OpenAI admitted the error and apologized.

What OpenAI says

OpenAI engineer Tarun Gogineni stated that reverse engineering with Codex is not prohibited. According to him, most of the blocks are erroneous, and the company is working on a fix.

The problem here is not a formal ban but the protection mechanism itself. Decompilation, disassembly, and protocol analysis look the same no matter who does them. A person cleaning ads off their own TV and someone hunting for a vulnerability to attack use the same techniques. Intent can't be extracted from a request, so a filter tuned with a safety margin inevitably hits honest users.

Why this matters in practice

Researchers' main complaint is not the refusals as such, but the unpredictability. Information security consultant Matt Brown told 404 Media that he switched to GLM 5.3, a local open-weights model. He uses it, among other things, to investigate hidden Bluetooth commands of a gas cylinder sensor. The reason is not only the restrictions: the rules can change without warning, and a tool that works today may become unusable tomorrow.

Similar scenarios were also covered on Habr. In October, an article was published about the "ChistyTV" project, which disables ads, telemetry, and unneeded apps on Android TV. For analyzing unknown firmware, the author recommends Claude Code or Codex, which are exactly the tasks that may now run into restrictions.

What to do about it

• Don't make a single cloud agent a critical link in any process that involves reverse engineering. Keep a fallback: a local open-weights model.
• For research work, it is better to set up a separate account. That way a false block won't take away your main work tool.
• If you build a product on an API, factor in the risk of false moderation triggers and think through an appeals path for users.
• Document the context of your tasks: that the device is yours and what the goal is. It will help when contacting support.

Limits of the story

Everything rests on user complaints and the comment of a single OpenAI engineer. There is no data on how many people are affected. Anthropic's position is not given in the retelling, so it is unclear whether this is a deliberate tightening of policy or a side effect of safety settings. Local models are not a panacea either: in quality they don't always catch up with cloud ones, and setting them up takes time.

The bottom line is simple. A paid AI service may refuse to analyze software on a device that belongs to you. And if a false block triggers, you also risk losing access to your account. While vendors tune their filters, it makes sense to have a fallback path that doesn't depend on someone else's rules.