Cheap AI-Assisted Reverse Engineering and What It Will Do

What it is

A October 2026 arXiv paper (cs.CY) by Ilan Mandel, Madiha Zahrah Choksi, Joey Castillo, Sebastian Benthall and James Grimmelmann (Cornell Tech, Columbia, NYU). It argues that the law on reverse engineering (RE) has not changed, but its practical effect has. Law permits RE largely because RE was slow and needed rare expertise. Coding agents, often $20 per month or less, remove that filter. The authors say this tilts the balance toward reverse engineers and that no policy chose this outcome.

Evidence the paper cites on capability

• SRE-Bench (unseen binaries, standard RE tools): GPT-5.5 fully solved 4% of instances. Four months later GPT-6 Astra solved 56.9% at about $13.50 per binary, even after refusing some tasks. Without safeguards it solved 88.0% on the first attempt and 99.2% within four.
• Open-weight models: NIST's Center for AI Standards and Innovation found Z.ai's GLM-5.3 trailed the U.S. frontier on cyber benchmarks by four months. On ExploitBench it developed 50 of 410 exploits, against 56 for Claude Mythos Preview. It complied 64% of the time when given a plausible cover story such as posing as a red-team agent. An abliterated version complied with every request.
• Refusals as a weak control: Eric Pardee spent months removing a reboot sequence from an Amazon Fire tablet they owned. After Claude stopped, they switched to Kimi K3, GLM-5.2 and GLM-5.3 and gained root. The paper notes this removal falls within current § 1201 exemptions, so the refusals blocked lawful work.
• Anecdotes: Schlarp reverse engineered consumer electronics with Claude running for a few hours per device. PhotoCraft (October 2026) reimplements Photoshop in Rust "from public specs and observed behavior only". Wine took 15 years to reach version 1.0 in 2008, which the paper uses as the old cost baseline.

Where the legal strain falls

• Copyright owners lose lead time as knock-offs get cheaper.
• Vendors of tethered goods face owner circumvention.
• Security researchers, repairers and interoperability developers still carry the risk under § 1201 and the CFAA. Even where an exemption allows circumvention, the tools stay unlawful to distribute.
• General-purpose agents can now match special-purpose circumvention tools such as tuya-convert for many uses.
• Contract terms assume human intent and consent.
• Without training-data disclosure, an AI "clean room" may be "at least a little dirty".

Four possible responses

flowchart TD
    A[AI makes RE cheap] --> B[Inertia: leave law unchanged]
    A --> C[Restore old balance: tighten laws]
    A --> D[Embrace AI RE: loosen constraints]
    A --> E[Tie RE privilege to device ownership]

The authors endorse none of these. They aim to show the problem exists and to map the options.

Caveats

The fetched text ends partway through Section 3, so the detailed mapping to trade secret, copyright, anti-circumvention, computer misuse and contract law, and the trade-offs of each response, were not available. The paper gives no operational guidance for engineers. Its benchmark figures are cited from other sources, not run by the authors.