Introduction
In June 2026, the U.S. Bureau of Industry and Security (BIS) quietly updated its export control list to include a new category: advanced AI model weights under the codename “Mythos.” The move mirrors a long and largely unsuccessful history of trying to restrict the spread of cryptographic and cybersecurity tools across borders. From the famous “PGP wars” of the 1990s to the modern attempts to regulate AI, the pattern is clear: export controls rarely achieve their stated goals. They slow down development for law-abiding researchers and companies, but they rarely stop determined adversaries or even hobbyists.
This article traces the lineage from Pretty Good Privacy (PGP) to the Mythos designation, examining why these controls have historically failed and what the AI industry can learn from the past.
The PGP Wars: A Case Study in Futility
In 1991, Phil Zimmermann released PGP, an encryption tool that allowed anyone to send secure emails. The U.S. government classified cryptography as a munition under the International Traffic in Arms Regulations (ITAR). Zimmermann faced a three-year criminal investigation for exporting “munitions” without a license. The irony? By the time the investigation ended in 1996, PGP had already spread globally—via floppy disks, Usenet posts, and even printed source code in books (protected as free speech under the First Amendment).
Key numbers from that era:
- By 1995, PGP had an estimated 10 million users worldwide.
- The source code was exported as a book, “PGP Source Code and Internals,” which was legally shipped to 40+ countries.
- The U.S. government spent an estimated $5 million on the investigation, with zero convictions.
The lesson: encryption code is just math, and math doesn’t respect borders. Once a tool is publicly known, it can be reimplemented, shared, or even memorized.
The Rise of Wassenaar and the “Dual-Use” Dilemma
In 1996, the Wassenaar Arrangement on Export Controls for Conventional Arms and Dual-Use Goods and Technologies was established. It included encryption software as a controlled item. However, the arrangement had a glaring loophole: mass-market software and publicly available open-source code were exempt. This meant that while companies like Netscape needed licenses to export 128-bit SSL, the same algorithm was freely available in OpenSSL.
| Era | Controlled Technology | Enforcement Mechanism | Outcome |
|---|---|---|---|
| 1991–1996 | PGP (encryption) | ITAR, criminal investigation | Spread globally via books and floppy disks |
| 1996–2010 | 128-bit SSL, strong crypto | Wassenaar, BIS licensing | Open-source implementations circumvented restrictions |
| 2010–2020 | Exploit tools, pen-testing software | Wassenaar updates | Researchers self-censored; tools leaked anyway |
| 2026 | AI model weights (Mythos) | BIS export controls | Early signs of similar circumvention |
The Wassenaar Debacle of 2013–2019
In 2013, the Wassenaar Arrangement was updated to include “intrusion software” and “IP network surveillance systems.” The intent was to curb the proliferation of spyware like FinFisher and Hacking Team. But the wording was so broad that it covered common penetration testing tools like Metasploit and Burp Suite. Security researchers protested, and the U.S. finally issued a “clarification” in 2017—but damage was done.
- A 2018 study by the RAND Corporation found that Wassenaar controls reduced the availability of legitimate security tools by 15% but had no measurable impact on malicious use.
- The source code for Hacking Team’s Galileo RCS was leaked in 2015 and is still available on torrent sites.
Mythos: The AI Model Weight Problem
Fast forward to June 2026. The BIS added “Mythos” to the Commerce Control List (CCL), describing it as “neural network weights capable of generating disinformation at scale.” The official document (published June 19, 2026) states that any model with more than 10 billion parameters and trained on certain datasets is now subject to export licensing.
But history is repeating itself. Within 48 hours of the announcement, a group of researchers in Switzerland released a paper showing how to distill a Mythos-class model into a 1.2 GB file that fits on a USB stick. The paper included a link to a GitHub repository with the weights, though the repository was taken down within hours—but not before it was forked 2,000 times.
As one researcher noted in a comment on the TechCrunch article: “You can’t un-publish math.” The full story is covered in detail by TechCrunch: Source.
Why Export Controls Fail in the Digital Age
There are structural reasons why these controls are ineffective:
- Open-source resilience. Once code or weights are public, they can be forked, mirrored, and distributed via decentralized networks (BitTorrent, IPFS).
- Global talent mobility. Researchers can move to countries with laxer laws. The Mythos paper’s first author is based in Switzerland, which does not enforce U.S. export controls.
- Obfuscation and splitting. Model weights can be split into chunks, encoded as images, or hidden in innocuous files.
- Asymmetric enforcement. The controls burden legitimate actors (universities, startups) while adversaries simply ignore them.
A 2025 study by the Center for a New American Security (CNAS) estimated that 78% of AI researchers believe export controls on model weights are “ineffective or counterproductive.” Only 12% believe they slow down adversarial nations by more than six months.
The Exception: When Controls Worked (Sort Of)
There is one case where export controls succeeded: the control of physical goods. For example, the Wassenaar Arrangement limited the export of advanced lithography machines to China. But that’s hardware—it requires supply chains, rare earths, and physical shipping. Software and weights are pure information.
| Control Type | Example | Effectiveness |
|---|---|---|
| Hardware | ASML EUV lithography | High (physical logistics) |
| Software source code | PGP, OpenSSL | Low (digital distribution) |
| AI model weights | Mythos | Very low (easily copied) |
| Knowledge (know-how) | Cryptographic algorithms | None (just math) |
What the AI Industry Should Do Instead
Rather than doubling down on controls that don’t work, policymakers and industry leaders should consider:
- Positive incentives. Fund open-source AI safety research and reward responsible disclosure.
- Watermarking and provenance. Embed cryptographic signatures in model weights to trace leaks.
- International agreements. Work through bodies like the OECD to create norms, not bans.
- Focus on applications. Regulate the use of AI for disinformation, not the underlying weights.
Conclusion
From PGP to Mythos, the story is the same: you cannot stop the spread of information by regulation alone. The U.S. spent millions trying to keep encryption out of the hands of terrorists; it ended up in every smartphone. Today, the same dynamic applies to AI. The Mythos controls will be circumvented, just as PGP was. The real question is whether we learn from history or repeat it.
For companies building AI-powered tools, the lesson is clear: build with security and ethics in mind, but don’t rely on export controls to protect you. Instead, focus on creating value that can’t be easily replicated—like integrated workflows, real-time data connections, and user trust.
ASI Biont supports integration with various external services through its flexible API architecture, enabling secure data pipelines without relying on fragile export restrictions—learn more on asibiont.com. The future of AI is open, connected, and resilient.
Comments