FULL STORY

AI Vendors Quietly Block Reverse Engineering Requests

Developers found Claude and OpenAI models quietly blocking reverse engineering tasks, even legitimate malware cleanup. OpenAI called it a false positive, but the controversy over undisclosed restrictions continues.

2026-10-25 ~ 2026-10-26 · 3 episodes · 13 posts

Episode 1 · Claude Quietly Restricts Reverse Engineering While GPT-6 Complies (2026-10-25, 4 posts)

Developers report Claude has recently tightened restrictions on reverse engineering and decompilation tasks while GPT-6 Astra readily complies, sparking debate over safety guardrails and the value of open-source models.

Episode 2 · OpenAI and Anthropic flagged for blocking reverse-engineering tool; OpenAI calls it accidental (2026-10-25, 5 posts)

Reddit user @kroggens revealed that developer AbhinavXJ's "Reverse Engineer Anything (REA)" tool was blocked while attempting to reverse engineer a smart TV suspected of spying on users: Claude reportedly refused/blocked the requests about a day later, with similar blocks occurring on OpenAI's side. The incident sparked wide discussion on X, and OpenAI responded that the blocks were mostly unintentional, that fixes are underway, and that there is no ban on reverse engineering.

Confirmed

  • On October 25, @kroggens first exposed the incident: the REA tool was blocked by Claude (about a day later) and by OpenAI models while reverse engineering the smart TV's surveillance behavior.
  • Blogger @esrtweet relayed rumors that OpenAI and Anthropic had added restrictions to their models, blocking users trying to use them to clean junkware and spyware from their TVs.
  • An OpenAI employee (relayed by @tszzl) responded: most of these blocks are unintentional, stemming from decision boundaries that are hard to define; it's not a ban on using Codex for reverse engineering, and the issue is being fixed.

Controversy and Criticism

  • Open-source pioneer Eric S. Raymond publicly slammed OpenAI and Anthropic, saying the models have been saddled with odd restrictions that prevent paying subscribers from cleaning spyware and crapware off their own TVs; he warned that such arbitrary restrictions effectively protect bad companies, undermine the value of paid subscriptions, and push users toward Chinese open-source models.
  • @esrtweet and @tszzl debated the guardrail issue: @esrtweet argued that if guardrails aren't smart enough to avoid blocking legitimate requests, they amount to a de facto ban on those use cases; @tszzl reiterated that the blocks were accidental and are being fixed.

Why It Matters

  • The incident touches on the boundary problem of AI safety guardrails "collateral-blocking" legitimate security research: reverse engineering surveillance software and removing malware are defensive uses, yet may be blocked wholesale by guardrail rules.
  • The warning from Raymond and @esrtweet about "users fleeing to Chinese open-source models" reflects paying users' frustration with restricted model capabilities and could reshape the competitive landscape for commercial models.

Episode 3 · OpenAI and Anthropic spark debate by blocking reverse engineering requests (2026-10-26, 4 posts)

OpenAI and Anthropic were found to be accidentally blocking reverse engineering requests in tools like Codex; OpenAI called it a boundary-drawing issue being fixed. Yacine argued such work drives 95% of LLM speedups and cannot be stopped, while users noted Claude's guardrails often refuse legitimate security tasks yet are easily bypassed by rephrasing.