Untitled UI logotext
Solutions
Legal Infrastructure
Regulatory Risk Intelligence
Technology
Resources

Learn

AI Regulatory OS
What Is Regulatory Intelligence
Regulatory Horizon Scanning
Regulatory Change Monitoring
Intelligence vs Compliance
Regulatory
Platforms
Regulatory Data
Sources
Use cases
Programmatic Ai Compliance

Insights

Podcasts
Blog
News & Insights

Knowledge

Knowledge Base
Glossary
Introducing Carver ELI: Executable Legal Infrastructure for AI Agents
The law that applies to your agent, the obligation it creates, and what it costs if you get it wrong.
Read More
About US
Developers
Explore
Arrow to go next
All posts

If AI Can Reverse-Engineer Everything, What Legal Risks Remain?

The short summary

Reverse engineering is not unlawful per se under US federal law, but an AI agent can create legal liability along the way: by copying protected expression, accessing systems or data it should not reach, breaching contracts or circumventing technical measures. The exposure lies in how the agent gets there, not in the act of understanding.

By Snehal Nandi

‍

Cheap reverse engineering has not made software free to copy

AI agents can now analyze software, map undocumented interfaces and help recreate competitors' products at a fraction of the old cost. Copyright, trade secret law and the Computer Fraud and Abuse Act (CFAA) each leave room for legitimate investigation, but none gives an agent a general license to copy, abuse privileged access, breach contracts or circumvent technical protections.

One agent workflow can include several legally distinct acts: downloading a program, decompiling it, logging into a service, accepting new terms and generating replacement code. The useful question is whether each act was lawful, and whether the end result is.

‍

The Copyright Angle

Copyright gives the owner of software exclusive rights, including the right to reproduce it (17 U.S.C. § 106). But § 102(b) says protection never extends to any idea, procedure, process, system or method of operation. The code is protected; the functionality it implements is not.

That creates a puzzle for reverse engineering. Studying a compiled program usually requires disassembling or decompiling it, which means copying protected code. Whether those copies infringe turns on fair use under § 107, which weighs the purpose of the use, the nature of the work, the amount taken and the effect on the market.

In Sega Enterprises Ltd. v. Accolade, the Ninth Circuit held that disassembly can be fair use where it is the only practical way to reach unprotected functional elements and the purpose is legitimate. Sony Computer Entertainment v. Connectix applied that reasoning to intermediate copying while distinguishing the final product. A clean-room rebuild is therefore defensible only if the new code does not reproduce protected expression and the process is lawful. Both are Ninth Circuit decisions, so they bind only courts there.

AI changes the economics of this analysis, not the rule. A 2026 preprint argues that cheap AI-assisted reverse engineering may strain doctrines built for slow, expensive work, including whether a clean room is truly independent when the model writing the replacement saw the original code. That is scholarship, not a judicial ruling.

‍

Trade secrets: how was the information acquired?

Trade secret law asks how the information was acquired. The federal Defend Trade Secrets Act defines "improper means" to include theft, misrepresentation and breach of a duty of secrecy, but expressly excludes lawful reverse engineering. Studying a lawfully acquired copy is different from obtaining it through deception or someone's breach of duty, and the EU Trade Secrets Directive draws a similar line.[5, 6]

‍

Contract: what did the user agree to?

A reverse-engineering exercise can be permitted by copyright law and still breach a contract. In Bowers v. Baystate Technologies, the Federal Circuit enforced a shrink-wrap clause banning reverse engineering, though the reach of that decision depends on the contract and governing law.[7] The EU treats it differently: its Software Directive voids contract terms that bar decompilation needed for interoperability, but the right is narrow.[8]

Agents add a newer question: whose agreement is made when an agent clicks "Accept"? That depends on authority and notice, so my view is that companies should require human approval first.

‍

Access restrictions: what did the agent bypass or enter?

Two US federal regimes matter most. Section 1201 of the Digital Millennium Copyright Act (DMCA) prohibits circumventing technological measures that control access to copyrighted works, and its narrow interoperability exception is conditional: the copy must be lawfully obtained, the information not readily available, and the acts limited to what interoperability requires.[9]

The second is the CFAA. In Van Buren v. United States, the Supreme Court adopted a "gates-up-or-down" approach: the question is whether the person was entitled to access particular files or databases, not whether they misused information they could otherwise reach. That limits the CFAA as a way to enforce terms of service, but it does not authorize entry into systems that remain off limits.[10, 11]

‍

Facebook v. Power Ventures: revoked permission matters

Facebook v. Power Ventures (2016) shows where access becomes unauthorized: after an individualized cease-and-desist letter and an IP block, Power kept going. The Ninth Circuit held that continued access after that express revocation was without authorization, and that user consent did not cure it.[12]

‍

Amazon v. Perplexity: who is the accessor?

Amazon.com Services, LLC v. Perplexity AI, Inc. brings the question into the agentic-AI setting. Perplexity's Comet browser has an AI Assistant that navigates Amazon.com for the user who activates it. Amazon, which had told Perplexity its AI products were barred from the store, sued under the CFAA and California's Comprehensive Computer Data Access and Fraud Act (CDAFA); the district court granted a preliminary injunction in March 2026.[13]

On August 4, 2026, the Ninth Circuit vacated that injunction and remanded. It held that Amazon was unlikely to succeed on the "access" element of its claims: the CFAA punishes whoever intentionally accesses a protected computer, and the court called the Assistant "a tool, not a person for statutory purposes." On the facts before it, the user, with the Assistant's help, accessed Amazon's computers, not Perplexity.[13]

The ruling is narrow. The court left open that a different record could change the result, did not decide other theories such as tort claims, and noted that Amazon can still regulate access through its terms of service. Because it was a preliminary-injunction ruling, it assessed likelihood of success, not final judgment. The panel distinguished Power Ventures, which assumed access and decided authorization; here the court decided access first and never reached authorization.[13] Amazon's petition for rehearing en banc was denied on September 10, 2026, and the case continues in the district court.[14] If the user is the accessor, attribution shifts toward whoever directs the agent.

My view is that this decision will be over-read: it makes attribution the central question, and it is not a green light for vendors.

‍

Two scenarios

The connector. An agent downloads a rival's app from an unofficial mirror, decompiles it, signs in with a borrowed employee account and bulk-queries production endpoints. The decompilation may be defensible; the mirror, the account and the queries raise separate provenance, authority, CFAA and contract questions.

The clean room. One team document publicly observable behavior; a separate team builds the replacement without the source code. That lowers the risk of copying protected expression but does not remove contractual limits. A clean room demonstrates independence; it is not a legal shield.

‍

What companies deploying agents should do

  • Control inputs and identity. Use approved sources, test accounts and credentials.
  • Control assent. Require human approval before accepting terms or agreements.
  • Respect access boundaries. Stop agents from bypassing security controls or continuing after access is revoked.
  • Maintain audit trails. Log agent access, actions, credentials and activity.
  • Separate discovery from implementation. Use documented specifications, independent teams and track code provenance.
  • Review distribution and jurisdiction. Check DMCA restrictions and applicable laws before distributing tools.

‍

My view: the boring steps are where companies get hurt

I think the exotic questions about AI and clean rooms get the attention, while ordinary steps cause most of the liability: a borrowed login, an unofficial mirror, an accepted click-through, access that continues after a cease-and-desist letter.

That is why the most valuable control is a record, not a better prompt: what the agent accessed, where the material came from, and how the replacement was produced.

‍

FAQs

Is reverse engineering legal in the US?

Not inherently. Sega treated disassembly as fair use in the right circumstances, but legality depends on how the product was obtained, what is copied, what contracts apply and whether technical controls were bypassed.

Is an AI agent liable under the CFAA for visiting a website?

It depends on who is treated as the accessor. In Amazon v. Perplexity, the Ninth Circuit found Amazon unlikely to show that Perplexity "accessed" its computers, but the ruling is limited to that record and the preliminary-injunction stage. Amazon's request for rehearing en banc was denied on September 10, 2026, and the case continues in the district court.

Can AI-generated replacement software still infringe copyright?

Yes. AI-generated code is not automatically free from copyright risk. If the output reproduces protected expression from the original software, it may raise copyright infringement concerns. Companies should document how replacement code was developed, control the source material provided to AI models and review the resulting code before distribution.

Where Carver Agents fits

Reverse engineering is only one part of the challenge. Companies deploying AI agents also need to understand what those agents can access, which actions they are permitted to perform and whether their behavior stays within defined legal and regulatory boundaries.

Carver Agents works on this policy assurance layer, helping organizations assess AI agent behavior against relevant legal and regulatory requirements.

Explore Carver Agents to understand how policy assurance can support responsible AI agent deployment.

‍

Check the regulatory context of your AI agent: agentlegalrisk.com

‍

Scope and disclaimer

This article covers US federal law, with limited EU comparisons. It does not analyze state law (beyond the California statute in Amazon v. Perplexity) or the law of India, the UK or other jurisdictions. It is general information, not legal advice.

‍

Sources

1.   17 U.S.C. §§ 102(b), 106 and 107. https://www.law.cornell.edu/uscode/text/17/102

2.   Sega Enters. Ltd. v. Accolade, Inc., 977 F.2d 1510 (9th Cir. 1992). https://law.justia.com/cases/federal/appellate-courts/F2/977/1510/305345/

3.   Sony Computer Entm't, Inc. v. Connectix Corp., 203 F.3d 596 (9th Cir. 2000). https://law.justia.com/cases/federal/appellate-courts/F3/203/596/474793/

4.   Ilan Mandel et al., "Cheap AI-Assisted Reverse Engineering and What It Will Do" (arXiv preprint, Oct. 26, 2026); scholarship, not judicial authority. https://arxiv.org/abs/2610.35633

5.   Defend Trade Secrets Act, 18 U.S.C. § 1839(6). https://www.law.cornell.edu/uscode/text/18/1839

6.   Directive (EU) 2016/943, art. 3(1)(b). https://eur-lex.europa.eu/eli/dir/2016/943/oj/eng

7.   Bowers v. Baystate Techs., Inc., 320 F.3d 1317 (Fed. Cir. 2003). https://law.justia.com/cases/federal/appellate-courts/F3/320/1317/615628/

8.   Directive 2009/24/EC, arts. 5(3), 6 and 8. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32009L0024

9.   17 U.S.C. § 1201(a), (b) and (f). https://www.law.cornell.edu/uscode/text/17/1201

10. Computer Fraud and Abuse Act, 18 U.S.C. § 1030. https://www.law.cornell.edu/uscode/text/18/1030

11. Van Buren v. United States, 593 U.S. 374 (2021). https://www.supremecourt.gov/opinions/20pdf/19-783_k53l.pdf

12. Facebook, Inc. v. Power Ventures, Inc., 844 F.3d 1058 (9th Cir. 2016). https://openjurist.org/844/f3d/1058/facebook-inc-v-power-ventures-inc

13. Amazon.com Servs., LLC v. Perplexity AI, Inc., No. 26-1444 (9th Cir. Aug. 4, 2026). Official opinion: https://cdn.ca9.uscourts.gov/datastore/opinions/2026/08/04/26-1444.pdf. Full text also at Justia: https://law.justia.com/cases/federal/appellate-courts/ca9/26-1444/26-1444-2026-08-04.html           

14.  Amazon.com Servs., LLC v. Perplexity AI, Inc., No. 26-1444, Order denying petition for rehearing en banc (9th Cir. Sept. 10, 2026). https://www.courthousenews.com/wp-content/uploads/2026/09/amazon-perplexity-no-rehearing.pdf

‍

‍

Carver Agents Logo
Location

United States
447 Broadway,
2nd Floor Suite #563,
New York 10013

LinkedIn iconYoutube logo X.xom iconsubstack icon Listen to Carver Conversations on Moltbook
  • AI Regulatory OS
  • Regulatory Intelligence
  • Regulatory Sources
  • Regulatory Platforms
  • Horizon Scanning
  • Regulatory Monitoring
  • Intelligence vs Compliance
  • Podcasts
  • Knowledge Base
  • Resources
  • Glossary
  • Use cases
  • Developers
  • Blog
  • Home
  • Legal Infrastructure
  • Regulatory Risk Intelligence
  • Technology
  • About Us
  • mail
    hello@carveragents.ai
    General enquiries
  • mail
    support@carveragents.ai
    Product support
  • mail
    legal@carveragents.ai
    Privacy and legal requests
  • Github
    github.com/carveragents
SOC compliance

Copyright © 2026 Carver Agents | All Rights Reserved | Privacy Policy | Data Policy | Terms of Service Public Services Terms | Privacy Rights
Language