Generative AI makes reverse engineering cheap and fast, threatening existing IP laws
The claim
Generative AI dramatically reduces the cost and difficulty of reverse engineering, undermining current legal protections
Published as: “Cheap AI-Assisted Reverse Engineering and What It Will Do”
So what
The ease of AI-driven RE is likely to force vendors to shift more software functionality to the cloud, reducing user control and reshaping the balance of IP incentives.
The five-minute version
AI-assisted reverse engineering makes RE dramatically faster, cheaper, and accessible to novices, threatening the assumptions of IP law that RE is costly and expertise-limited. The paper documents recent LLM capabilities (e.g., GPT-6 solving 56.9% of RE benchmarks) and argues this will erode legal protections and push vendors to move functionality to the cloud. Evidence is largely qualitative and based on recent model releases and anecdotal case studies; no systematic empirical evaluation is presented.
What they found 2
- GPT-6 Astra solved 56.9% of RE benchmark instances at about $13.50 per binary.
“GPT-6 Astra solved 56.9% overall at about $13.50 per binary, even after refusing some tasks.”
- Claude can run autonomously on a device for hours with minimal user interaction.
“Claude ran for a few hours with each device plugged into his computer.”
What undercuts it the paper’s own results
- Safeguards offer limited protection, and open-weight models diffuse these capabilities further.
- Vendors will lock on-device functionality behind technical barriers, attempting to slow AI-assisted RE.
Stated limits 1
- We do not endorse or reject any of these proposals; we aim to show that the crisis exists and to map the range of possible responses.
What’s new here 4
- Maps recent large-language-model RE capabilities onto existing IP, copyright, trade-secret, and computer-misuse law.
- Identifies four distinct policy response pathways that have not been systematically outlined before.
- Provides up-to-date performance numbers for AI models on RE benchmarks (e.g., GPT-6 56.9% success).
- Highlights the diffusion of open-weight models as a factor accelerating the legal crisis.
Artifacts & methods 4
- large language model agents (e.g., Claude, GPT-6 Astra)primitive
- SRE-Benchdataset
- ExploitBenchdataset
- Open-weight models (GLM-5.3, Kimi K3)system