Has Anthropic poisoned its own well by arguing its new Mythos model was so dangerous it couldn’t be released publicly?
After Amazon researchers discovered a jailbreak of Fable 5 (Mythos with guardrails), the U.S. government banned use of both Mythos and Fable 5, exactly the kind of government scrutiny that could disrupt Anthropic’s business most.
Amazon CEO Andy Jassy was reportedly the source of security concerns that led the U.S. government to force Anthropic to shut down Fable 5 and Mythos 5 for every customer, also imposing an export control ban on both models.
Oddly enough, Amazon is one of Anthropic’s largest investors, having put in billions and receiving a $100 billion cloud spending commitment in return.
Anthropic says the jailbreak technique surfaced “a small number of previously known, minor vulnerabilities.”
It called the government’s response disproportionate and said the capabilities causing concern are already available in other publicly accessible models. The shutdown affected every customer globally because Anthropic cannot filter foreign nationals from U.S. users in real time.
For Anthropic, the immediate question is how quickly it can restore access. And that will happen.
For the broader industry, the precedent is what matters. Anthropic has been lobbying for more regulatory power over AI. It just found out it might get what it wants.
Anthropic's own CEO publicly called for the government to have authority to block dangerous model releases. The government then used such calls against Anthropic two days later.
Optimists might argue that government restrictions on dual-use technologies might slow down technology diffusion, but cannot halt it.
The general pattern across technology history is this: governments impose controls with genuine security rationale, those controls prove partially effective in the short term, generate substantial unintended costs, and are eventually relaxed or circumvented.
The regulated party frequently bears costs that accrue to unregulated competitors.
Technology | Regulation | Period | Short-Term Effect | Long-Term Outcome |
Cryptography, Encryption (PGP, SSL) | U.S. export ban; treated as a munition under ITAR; Clipper Chip mandate attempt | 1975 to 2000 | Phil Zimmermann prosecuted 3 years for posting PGP online; RSA export restricted New America · Brookings | Controls collapsed when it became obvious software could not be bottled; Clinton-era deregulation followed. In subsequent decades, virtually all predictions about encryption's economic benefits were borne out — SSL, SSH, electronic banking, e-filing, and VPNs depended on exactly the strong encryption the government tried to restrict. |
Cryptography | Export controls treated software as equivalent to munitions | 1990s | The Zimmermann case dragged on three years before being dropped, and export control laws were eventually rewritten after it became obvious that software couldn't be contained like rocket motors. Crowe LLP · Wassenaar/arxiv | The EU abolished cryptographic export controls within the union in 2000, a decision subsequently adopted by the Clinton administration; more recent trends have involved further relaxation rather than tighter controls. |
CoCom, Cold War Computing | Coordinating Committee for Multilateral Export Controls: Western embargo on technology exports to Soviet Bloc | 1949–1994 | In the period 1951–1967 CoCom performed reasonably well despite its known limits. After that, enforcement eroded. ChinaTalk/Ottinger · Texas Nat'l Security Review | CoCom's effectiveness was reduced throughout its history by overt non-compliance, differences between member nations, the secretive regime, and the financial incentives of tech exporters. CoCom did not prevent the USSR from accessing key technologies; the size of the control regime expanded over time, weakening enforcement and encouraging defection. Disbanded 1994. |
Toshiba-Kongsberg Scandal | Sanctions on Toshiba for selling machine tools enabling quieter Soviet submarine propellers | 1982–1987 | Washington imposed sanctions on Toshiba worth $30 billion amid congressional claims the breach jeopardized U.S. national security. AFSA | The $30 billion estimate was primarily based on a hyperbolic claim that the entire U.S. nuclear submarine fleet would need to be replaced — the actual magnitude of the damage remained unclear. Sanctions were eventually relaxed; Toshiba survived. |
Nuclear Technology | Nuclear Non-Proliferation Treaty (NPT); Nuclear Suppliers Group; export controls on enrichment technology | 1968–present | By the 1950s, it became clear that pure denial had neither stopped the Soviet Union nor the UK from acquiring nuclear weapons. The 1954 revision of the Atomic Energy Act reflected a shift from prevention-through-denial to influence-through-cooperation. CNAS · AADS | U.S. dominance in commercial nuclear energy in the 1970s allowed the U.S. to keep Taiwan and South Korea from developing their own nuclear facilities — a genuine success. But Pakistan, India, North Korea, and Israel all acquired nuclear capability despite controls. NPT remains the most durable tech-control regime, though partial. |
Semiconductors, AI Chips (China) | BIS October 2022 controls on NVIDIA A100/H100 and manufacturing equipment; tightened 2023, 2024, 2025 | 2022–present | Implementation significantly disrupted China's semiconductor ecosystem, causing price spikes for some device types and forcing workforce reductions. CSIS/HSToday · CSIS | Loopholes, alternative approaches, and unforeseen outcomes have diminished long-term efficacy — the very restrictions designed to hinder Beijing may instead be accelerating China's domestic progress. Companies impacted by the October 2022 controls, on average, even outperformed comparable unaffected companies, exhibiting higher increases in R&D spending and patent filings — a counterintuitive outcome driven by AI chip demand. |
Semiconductor Controls | EAR controls tightened repeatedly; allied coordination with Netherlands, Japan | 2022–present | According to research, versions of Intel's Xeon Gold and NVIDIA GeForce RTX chips can be bought on Taobao; students at Tsinghua report being able to "easily circumvent" restrictions on U.S. EDA software. Texas NSR | CoCom did not prevent the USSR from accessing key technologies; the current regime is similarly porous, and China is a more adept target. Long-term verdict still open. |
AI Models Fable 5 / Mythos 5 | U.S. Commerce Dept. emergency export control directive; all foreign nationals barred | June 2026 | Anthropic had to cut off access to both models for all customers worldwide because it could not rapidly implement nationality-based filtering. Anthropic disputed the severity of the jailbreak, arguing many of the same vulnerabilities could be discovered using other publicly available models. Fortune · Axios · State of Surveillance | Too recent to assess long-term outcome. The next letter may go to a different vendor; the policy tools used here will be available to every future administration. Watch for restored access accompanied by new conditions such as nationality verification, tighter pre-deployment review, or federal observer access. |
Several durable findings emerge from this body of evidence:
Unilateral controls erode without multilateral coordination
The control regime tends to expand and become self-defeating
Denial often accelerates indigenous capability in the target
Software and information resist containment
The regulated party bears costs; competitors do not.
Rival Sam Altman, of OpenAI, says Anthropic’s fear-based marketing now has come back to bite Anthropic.
No comments:
Post a Comment