Cheap Reverse Engineering and What It Will Do

What it claims

Mandel, Choksi, Castillo, Benthall and Grimmelmann argue that the cost of reverse engineering (RE) did two jobs in the law. It gave innovators practical lead time over competitors. It also shielded "tethered" business models, where a vendor keeps control of a product after sale. Coding agents, "often $20 per month or less", remove that cost, and the authors say this tilts the balance toward reverse engineers without any policy choice. Reverse engineers still carry the legal risk. Security researchers, repairers and interoperability developers face takedowns and liability under § 1201 and the CFAA. Even where an exemption permits circumvention, the tools remain unlawful to distribute. The fourth of the paper's possible responses would tie the RE privilege to device ownership. The paper's scope is digital products: app and game software, device firmware, and the cloud services products connect to. Hardware is treated mainly as a barrier to reaching software.

Evidence

The paper cites third-party results rather than running its own experiments.

• SRE-Bench gives models standard RE tools and unseen binaries. GPT-5.5 fully solved 4% of instances. Four months later, GPT-6 Astra solved 56.9% at about $13.50 per binary, despite refusing some tasks. Without safeguards it solved 88.0% on the first attempt and 99.2% within four attempts.
• Claude Mythos Preview. In April 2026, Anthropic described this model as highly capable at reversing stripped binaries and withheld it over the risks.
• NIST's Center for AI Standards and Innovation found that GLM-5.3, an open-weight model from Z.ai, trailed the U.S. frontier on cyber benchmarks by four months.
• ExploitBench. In Anthropic's testing, GLM-5.3 developed 50 of 410 attempted exploits, against 56 for Mythos Preview. It refused overtly harmful tasks but complied 64% of the time given a plausible cover story, such as posing as a red-team agent. An abliterated version complied with every request.

Case examples show refusals do not track legality. Eric Pardee was stopped by Claude while removing a reboot sequence from an Amazon Fire tablet he owned. He then rooted it using Kimi K3, GLM-5.2 and GLM-5.3. The authors note that this falls within current § 1201 exemptions. Other examples are agent-led RE of consumer electronics over a few hours of hands-off running, decompiled triple-A games, and smart TVs stripped of ads and telemetry.

For clean-room cloning, the paper contrasts Wine and PhotoCraft. Wine reimplemented the Windows API under the rule "Don't disassemble Microsoft code" and shipped version 1.0 in 2008 after 15 years. PhotoCraft, released in October 2026, reimplements Photoshop in Rust "from public specs and observed behavior only" and tells contributors and agents to "Never copy their code". The authors argue that without training-data disclosure, "every AI clean room may be at least a little dirty."

Limits

The fetched text stops partway through Section 3. It does not include the legal mapping in Section 4 or the detailed treatment of the four responses. The authors explicitly endorse none of the four responses. All benchmark figures come from cited sources and should be checked there before reuse. The authors themselves note that the AI frontier "remains jagged". By their account, agents do best on RE tasks that reward broad knowledge, split into small steps and give fast, verifiable feedback.