OpenAI’s AI coding tool Codex does not prohibit reverse engineering. That is according to OpenAI researcher Roon, who addressed the issue directly on X.
Roon also said that cases where Codex refused to help users remove software from their own devices were accidental, and that those blocks are being fixed.
What Roon actually said
The clarification landed on X on October 26, 2026, at around 00:43 UTC. Roon’s message had two parts.
First, Codex has no policy against reverse engineering. Second, the refusals users had run into around stripping software off devices were not intentional, and OpenAI is working to correct them.
Users flagged instances where the tool blocked requests tied to removing malware or cleaning up devices.
The criticism that set it off
Roon’s comments did not arrive in a vacuum. On October 25, 2026, open-source advocate Eric S. Raymond publicly called out both OpenAI and Anthropic.
Raymond is a well-known name in open-source circles, so his criticism carries weight with exactly the developer audience these AI coding tools are trying to win.
According to the research summary of the episode, the restrictions on Codex were unintentional and stemmed from safety filters that were drawn too broadly. In other words, guardrails meant to stop genuinely harmful requests ended up catching legitimate ones too.
The complaints also carried a competitive edge. Users raised concerns that overly restrictive behavior could push customers toward rival AI options from China.
This is also not a brand-new grievance. Complaints about Codex blocking legitimate reverse engineering first surfaced as early as June 2026.
Background: Codex’s rise and its growing pains
OpenAI launched Codex in April 2025. Codex had more than 2 million weekly active users by March 2026.
The product comes in two flavors. There is an open-source command-line interface released under the Apache 2.0 license, alongside a proprietary desktop version.
Yet Codex has faced usability complaints tied to cautious safety filters. The reverse engineering refusals are one of the more visible examples of that tension.
What this means for users and the AI coding race
For developers, the immediate takeaway is practical. Reverse engineering is not off limits in Codex, per Roon, and the device software blocks should go away as fixes roll out.
For OpenAI, the clarification could help limit frustration among existing users. The research summary suggests the move may help the company hold onto its customer base as interest grows in less restricted AI alternatives.
The competitive pressure is real. AI coding assistance is a crowded market, and Raymond aimed his criticism at Anthropic as well as OpenAI.
The concern about Chinese alternatives adds a geopolitical wrinkle. If US-based tools are perceived as more restrictive, some developers may be willing to trade other considerations for a model that simply does what they ask.
OpenAI’s framing, that the blocks were accidental rather than a policy change, is notable. It suggests the company is fixing a calibration problem rather than rewriting its rules.
The episode also highlights how much these products now depend on public pressure to correct course. A complaint first raised in June drew a clear answer in October, shortly after a high-profile critic spoke up.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.