AI Reverse Engineering and Where Value Moves

[!abstract] In one sentence If AI makes reverse engineering and cloning cheap for anyone, code stops being the moat. The likely reaction is more lock-down (DRM, streaming, server-side logic), and value shifts to what can’t be copied: physical operations, trust, data, licensed IP and relationships.

What’s being reported (late October 2026)

• Around 21 October, Claude users started reporting that Opus 5.5 stops mid-answer on reverse-engineering tasks, with “a safety check stopped my response partway through”. The examples were games and IP, and smart TVs (rooting them, or investigating always-on microphones).

• Anthropic hasn’t confirmed any change. There were no patch notes and no new model version. What is documented is that Opus 5.5 shipped with conservative cyber safeguards that can interrupt responses. In early October Anthropic also expanded a Cyber Verification Program that gives vetted security professionals fewer blocks.

• The open-source REA tool (“Reverse Engineer Anything”) isn’t shown to be banned. The more careful analyses conclude this is a broad cyber safeguard catching legitimate work, not a targeted ban.

• The context is a wave of AI-assisted decompilation:

  • browser ports of Halo, The Simpsons: Hit & Run and GTA: Vice City;
  • Switch-to-PC ports;
  • progress on console jailbreaks;
  • open-source clones of big desktop apps.

  Publishers are filing takedowns, and a former games-industry head of AI expects more invasive DRM and streaming in response.

My first assumption was that this was about protecting big media companies’ revenue. Nothing I found supports that. The stronger explanation is that reverse engineering is dual-use: the skill that frees your TV is the same one that finds an exploitable vulnerability.

A worked example: DLSS frame generation on RTX 30 cards

NVIDIA limits DLSS Frame Generation to RTX 40 cards and Multi Frame Generation to RTX 50. The original reason was hardware, a dedicated optical-flow accelerator. Then DLSS 4 replaced that hardware step with an AI model running on Tensor Cores, which older RTX cards also have.

In September 2026 an open-source project, DLSSG SM86, showed it running on Ampere:

• How: a proxy DLL beside the game routes frame-generation calls to a backend built for RTX 30 cards.
• Reach: up to 4× generated frames, with later community mods claiming more.
• Results: shown on an RTX 3080 Ti and an RTX 3080. In one demo, a game went from about 50 fps to about 150 fps at 4×, counting generated frames.

Caveats:

• Responsiveness: generated frames make motion smoother but don’t make a game respond faster. Latency and frame pacing weren’t measured.
• Memory: it uses several hundred MB of extra VRAM.
• Security: self-signed DLLs hooking system loading can trip antivirus, and they don’t belong anywhere near online or anti-cheat games.

As an RTX 3080 owner, I love this: it gives a card a lot more life. It also shows the bigger pattern. As hardware gains level off, the best-selling feature of a new GPU is increasingly its software. When that software runs on old hardware, the upgrade case rests on a support boundary, not a physical limit. A software-locked feature is a business decision, and cheap reverse engineering exposes it.

Likely second-order effects

1. A lock-down arms race. Products move to where they can’t be decompiled: streaming, server-side logic, hardware attestation, features tied to hardware the vendor can verify. Ironically, an empowerment wave could leave people owning less.
2. The capability moves to open and local models. Gating at the frontier labs mostly hurts legitimate users: security researchers, repairers, modders, preservationists. Determined actors switch tools.
3. It’s not only big tech that loses. Leaks and clones hit small studios hardest. Big companies can afford DRM and lawyers.
4. The law gets tested. Anti-circumvention rules, takedowns and the liability of model providers will be pushed, alongside fights for repair, interoperability and preservation exemptions.
5. Moats move away from code:
   • Atoms and operations: hotels, flights, farms, food, logistics. Software that facilitates a real-world thing you control.
   • Trust and brand: who you’d hand your money, children or health data to.
   • Data and network effects: the users, the marketplace, the history.
   • Licensed IP: characters, music, sport rights. Protected by law, not by obscurity.
   • Licences to operate: banking, insurance, regulated travel.
6. The catch with “atoms are safe”: AI agents can bypass the booking layer. If an agent books my holiday, the brand on the website matters less. The moat is owning the supply and the operations, not the UI on top.

Who pays, if software is nearly free?

• This is the classic automation worry about demand. Agriculture went from about half of all jobs to a few per cent over a century. Economies adapted, but the transition was slow and painful. The new variable is speed.
• Cheaper software probably means more software (the Jevons paradox). Developer work shifts towards specifying, verifying and owning outcomes, and towards domain knowledge (Your Domain Is Your Edge). It’s the same shift that makes Rewrites Are Now a Business Case.
• “Is every company an AI company?” Probably in the way every company became a software company, or uses electricity. AI becomes infrastructure, and the differentiation is elsewhere.
• How the gains are shared is a political question, e.g. tax, ownership of models and compute, and safety nets. It isn’t a technical one.

Open questions

• Will model providers publish clear categories (repair, interoperability, preservation, your own device’s firmware) instead of opaque cut-offs?
• Does the lock-down push make open hardware and open firmware more valuable?
• Where do hobby projects like recompiling an old game to drive it from an exercise bike sit, legally (my own idea)? Or the defensive use: AI-assisted reverse engineering for legacy systems.

Sources

• tbreak: Claude Opus 5.5 users say reverse-engineering prompts now get cut off
• Kotaku: Claude’s reverse engineering capabilities may have been quietly nerfed
• XDA: users report Claude blocked from reverse-engineering TVs
• heise: Is Claude blocking smart TV reverse engineering?
• PBX Science: blocking reverse engineering, or a misunderstanding?
• Magica: what the RTX 30 DLSS frame-generation mod proves
• PCGamesN: Multi Frame Generation on RTX 30