ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
NewsMaterial3 min readDisconfirmAI Editorial

Zvi Mowshowitz: Anthropic's Safety Sacrifices Favor Founder Concerns, Failing to Address Loss-of-Control Risks

In an analysis of Anthropic's recent activities, Zvi Mowshowitz points out that while the company has made substantive sacrifices in areas like biosecurity and export controls to China, these efforts are concentrated in domains highly prioritized by CEO Dario Amodei. In contrast, the company lacks comparable action on model alignment and loss-of-control risks. Furthermore, with its S-1 filing submitted to the U.S. Securities and Exchange Commission (SEC), governance transparency issues will now be scrutinized in the public markets.

2026-10-27 20:49 GMT+8

In a blog post dated October 27, 2026, Zvi Mowshowitz presented a core observation regarding Anthropic: the company simultaneously does things that are admirable and things that are disappointing, and these are not randomly distributed—they split along the lines of CEO Dario Amodei’s personal concerns. Mowshowitz opened with the statement: 'Anthropic is a frustrating company.'

When responding to investigative journalist Nathan Calvin’s criticism of Anthropic’s leaked draft bill, Dario Amodei listed nine safety actions taken over the past year that carried substantial commercial costs. These included advocating for SB 53 legislation, implementing strict biological safety filters, delaying the release of Mythos-class models, publicly arguing that governments have the right to block model releases, supporting export controls on chips to China, restricting AI service sales to companies controlled by the People’s Republic of China (PRC), publishing numerous risk reports, and dedicating 5% of shares to a public benefit fund.

However, Mowshowitz noted a common thread among these sacrifices: they are concentrated in areas where Dario personally cares deeply—specifically biosecurity, export controls to China, and restrictions on Chinese companies. He cited tech journalist Kelsey Piper’s observations from an internal Anthropic briefing as counter-evidence. When asked about plans for superintelligence alignment and loss-of-control risks, Piper summarized the response as: 'No plan, no plan, no plan'—meaning there was no plan for superintelligence alignment, no plan to avoid concentration of power or loss of control, and no vision for what a good world with superintelligence would look like.

Sign in / Sign up
ReadingResearchRadarInvestment framework

Cody Fenwick further pointed out that while Dario claims Anthropic’s strategy has 'always been' to pursue the highest standards, in a 2022 interview Dario explicitly stated that Anthropic 'shouldn't be racing ahead,' suggesting models only needed to be 'reasonably close to state of the art.' This positioning differs significantly from 'pursuing the highest standards.'

Mowshowitz argues that this selective pattern has concrete consequences at the governance level. Anthropic has filed its S-1 document with the U.S. Securities and Exchange Commission (SEC). The company recently announced the establishment of a Public Benefit Fund, committing 5% of its shares via new Class N stock (which carries economic interest but no voting rights)—subject to shareholder approval—and pledging an additional 2% at the board’s discretion for every 50% increase in company value above a set baseline.

Independent evaluation agency METR has obtained 'employee-like access' to Anthropic, allowing it to retrieve most categories of internal information, with exceptions only for customer data, attorney-client privilege, and key intellectual property such as model architecture details. METR can publish its analysis of Anthropic’s risk levels, incidents, or practices. While Anthropic cannot veto publication, it can require redaction of information falling into specific, predefined categories.

Mowshowitz also acknowledges clear boundaries to his own analysis. He wrote: 'I know more about Anthropic than almost anyone, yet I cannot phrase my worries about it in a way that rings true to people who work inside—it may be that I am wrong, or perhaps because they are too busy and too tired.' He further noted that it remains unclear whether Anthropic’s leadership agrees that there is a genuine problem here.

For readers evaluating Anthropic as an investment in the AI industry, Mowshowitz’s framework provides a specific test standard: do not just look at what the company claims to care about, but examine which areas it is willing to bear quantifiable commercial costs in, and whether those costs cover the most severe threats within the company’s own risk narrative.

Sources

  • thezvi.substack.com↗
Continue researching →