AI Agents Speed Up Game and Software Decompilation, and CVEs Hit 67,000 in 2026
Decompilation means recovering a program's logic from a finished binary without its source code. It used to take a team years to do this for a game. With AI agents, it now takes days. That makes it possible to move old games to new platforms, but it also makes it easier to hunt for unknown holes. As of September 2026, the NVD database holds more than 67,000 CVE identifiers, compared with 48,185 for all of 2025. The link to AI is still a hypothesis, but closed code should now be considered readable.
AI-processed from Habr AI; edited by Hamidun News
Imagine you spent years developing a paid application and believed that a binary file reliably hid your logic. Machine code has no function names and no comments, and analyzing it by hand would take months. Now neural networks and AI agents can recover the understandable logic of such a program in a matter of days. This process is called decompilation. It revives abandoned games, but it also makes hunting for vulnerabilities nobody knows about easier. For developers and owners of commercial software, this is already a practical question, not a philosophical one.
What has changed
First, the terms. Decompilation does not restore the real source code. It reconstructs an approximate representation of how a program works, and from that logic you can build a working replica. So the phrase "forced open source" is just a metaphor: such code has no licenses, commit histories, or comments.
Previously the work was collective and slow. Classic examples are OpenGOAL, which reconstructs GOAL, the language the Jak and Daxter series is written in, and Ship of Harkinian, a reconstructed console game with source code and binaries published on GitHub. These projects took years. Now, as the review's author writes, such tasks take days with AI agents.
The most striking examples involve games. Old Flash games are being recompiled to C++, and the new version runs at 1440p and 165 frames per second. The AnyPS5 project aims to automatically convert any PlayStation 5 game into binaries for Linux and Windows. It is not finished, but some games, such as Animal Well, have already been ported. This is not an emulator but rather a compatibility layer. A Windows program can be rebuilt for Linux, Android, or WebAssembly and run directly in a browser.
The flip side: vulnerabilities and code theft
The same techniques work for attackers. If an attacker decompiles a program, they search it for 0-days, vulnerabilities that neither the authors nor the defenders know about, and can exploit them immediately. Linus's Law ("given enough eyeballs, all bugs are shallow") does not help here: the code is read against the author's will, and nobody is obliged to report what they find.
The numbers look alarming. CVEs are public identifiers of known vulnerabilities, collected in the NVD database. According to its data, 2026 will be the first year in which the number of entries exceeds 50,000. By September there were already more than 67,000, and year-end estimates range from 70,000 to 96,000. The record-setting 2025 ended at 48,185. The CISA KEV catalog, which lists vulnerabilities actually being exploited, added 249 entries in the incomplete 2026, more than in all of the previous year. According to Mandiant's M-Trends 2026 report, the average time from a vulnerability's disclosure to its exploitation has become negative: minus seven days. In other words, many attacks begin before public disclosure.
An important caveat: the author only writes that "some experts suppose" this growth is linked to neural networks and mass decompilation. The material offers no direct evidence. The growth in CVE numbers was observed before, and AI also speeds up the work of defenders.
Among the possible threats, the author names polymorphic LLM worms that rewrite their code for any OS, and hardware modding: extracting and analyzing the BIOS of smart home devices and embedded systems. Another danger is backdoors in proprietary software, although a supply-chain attack still requires access to the development infrastructure, updates, or signing. There is also legal risk: services such as Clean Room as a Service (for now a joke) are appearing, offering to rewrite free code and eliminate license obligations. A paid application can leak into the open or be copied by a competitor.
What to do in practice
The author expects companies to respond with tougher DRM. Familiar measures also work: compiler optimizations, inlining, symbol stripping, and obfuscation greatly complicate decompilation, although they do not stop it completely.
For teams, the conclusion is simple. Do not rely on "security through obscurity": assume the binary will be read, and do not keep secrets or keys inside it. Shorten the window between a patch's release and its installation, since attacks now precede disclosure. Use the same tools for defense: code audits, bug hunting, and incident analysis with LLMs. The real scale of neural networks' impact on security is still hard to assess, and the situation is changing fast. But the economics of code analysis have already shifted: what once cost months of work now costs days.
- Nov 21, 2026Nvidia Board Ousts Jensen Huang, Putting $12.9B Hugging Face Deal at Risk
- Nov 21, 2026Mistral Opens Singapore Hub for Asia, Plans to Grow to 200 Staff by 2027
- Nov 21, 2026AIruntime builds a business card website from an AI chat for 50–300 rubles in tokens
- Nov 21, 2026US Closes Loophole: China Cut Off From Remote Access to Nvidia Chips Abroad
Key Questions Answered
What has changed?
First, the terms. Decompilation does not restore the real source code. It reconstructs an approximate representation of how a program works, and from that logic you can build a working replica. So the phrase "forced open source" is just a metaphor: such code has no licenses, commit histories, or comments. Previously the work was collective and slow.
What are the key details on The flip side: vulnerabilities and code theft?
The same techniques work for attackers. If an attacker decompiles a program, they search it for 0-days, vulnerabilities that neither the authors nor the defenders know about, and can exploit them immediately. Linus's Law ("given enough eyeballs, all bugs are shallow") does not help here: the code is read against the author's will, and nobody is obliged to report what they find. The numbers look alarming. CVEs are public identifiers of known vulnerabilities, collected in the NVD database.
What to do in practice?
The author expects companies to respond with tougher DRM. Familiar measures also work: compiler optimizations, inlining, symbol stripping, and obfuscation greatly complicate decompilation, although they do not stop it completely. For teams, the conclusion is simple. Do not rely on "security through obscurity": assume the binary will be read, and do not keep secrets or keys inside it. Shorten the window between a patch's release and its installation, since attacks now precede disclosure.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.