A recent employee resignation and a cybersecurity incident initiated by artificial intelligence agents have highlighted some of the safety, governance and transparency concerns investors face when evaluating upcoming initial public offerings from Anthropic and OpenAI.
The two companies filed for IPOs earlier this year, which will be closely watched, given their status as purely generative AI companies. The filings followed a SpaceX IPO that included xAI folded into the larger company. Both companies submitted confidential securities filings planning those initial offerings to the Securities and Exchange Commission in June, leaving potential investors in the dark about various aspects of the planned stock offerings.
Since Anthropic and OpenAI both opted for confidential S-1 filings to the SEC, investors will have to make their own decisions on what trustworthy AI looks like before those become public. Once the companies’ plans for initial offerings become public, investors should watch what they have to say about the companies’ AI safety and governance structures, according to a report from Morningstar Sustainalytics released Thursday.
“Trustworthy AI [models] are increasingly turning from a purely ethical or academic question to a tangible question and a tangible impact that is measurable even at the financial level,” Kilian Theil, the report’s author and director of the methodology team for Sustainalytics, told ESG Dive.
An AI training researcher, Jacob Coxon, publicly resigned from Anthropic, announcing his departure in a series of posts on X, formerly Twitter. Coxon also previously worked in similar research roles at Anthropic and OpenAI, according to his posts and verified emails on peer review site OpenReview and Google Scholar. In his posts, Coxon claimed “neither company is acting responsibly.”
Among the concerns raised in his series of resignation posts, Coxon said that the people building OpenAI and Anthropic’s models “earnestly believe that it could kill us all by the end of the decade.” Anthropic Alignment Science Organization Lead Evan Hubinger quoted that post and said in his own post on X that members of the company “do earnestly believe AI could kill all humans.”
Anthropic had to initially pause the release of its latest model due to federal government safety concerns. Meanwhile, open-source AI platform HuggingFace revealed in July that autonomous OpenAI agents had hacked its system; OpenAI revealed last month that 1,200 agents went rogue in the lead up to the attack and around 700 agents ultimately participated.
“Neither structure is demonstrably safer; they place different verification burdens on the buyer,” Morningstar’s report said when comparing what’s publicly known about the risk and safety of each company.
Theil’s Sept. 10 report touches on what investors can learn from the way the two AI companies have publicly handled the regulatory and cyber incidents. The report examines Anthropic and OpenAI through the lenses of political and societal stewardship; integrity; risk and safety; accountability; transparency; and environmental stewardship.
Anthropic is widely expected to debut its IPO sometime late this year, with some reports suggesting an offering as early as late September or early October. Though OpenAI was initially expected to also debut its IPO on stock exchanges this year, it is now targeting a date in 2027 “or sooner,” its CFO Sarah Friar reportedly told employees last month. Valuations for both companies are approaching $1 trillion as of Q2 of this year, according to data from PitchBook.
The question is not about “which company is more trustworthy” or “do people trust these companies,” Theil said in an interview. While they each serve different customers — with Anthropic focusing on enterprise customers and OpenAI having a larger consumer and free user base — Theil said their adoption rate proves people trust the platforms.
“People trust these companies for convenience because the models are increasingly powerful, but that's probably for not the entirely right reasons,” Theil said.
Theil said there remain questions of trust around the companies’ governance, transparency and what data their models are trained on, along with questions about power availability, environmental and societal stewardship and the regulatory landscape “that are ultimately more interesting, and that also … will have an important and profound impact on both companies.”
In evaluating the governance structures of the companies, Theil expects the S-1 filings to eventually shed light on the governance structures and the “protective mechanisms” that hold the companies accountable to their mission statements.
“This is something that is important because it then allows [investors] to compare what the company publicly proclaims,” Theil said. “What are the actual mechanisms that allow it to do so even as board composition changes, even as shareholders and share composition change, how is the company to be trusted, to continue to uphold that statement.”
However, the report expects investors and the broader market to anchor their expectations for the companies around enterprise revenue quality; governance stability; regulatory compatibility and transparency around the cost of compute power.
“Markets appear to recognize the upside to trust more readily than the downside from governance, regulatory, and compliance risks,” the report said. “Where fundamentals are comparable, firms that demonstrate stronger governance, safety controls, transparency and regulatory preparedness may carry lower downside risk.”
