Jasmine Wang, an OpenAI alignment researcher, posted Wednesday evening that “it’s hard to overstate how dangerous speeding towards RSI is,” referring to recursive self-improvement, the process by which models get better at improving their own capabilities. Her Anthropic counterpart Anna Wang made the same point hours later: “there is not yet a viable scientific plan to solve risks from recursively self-improving AI.” Two researchers, two labs, one week, the same technical worry stated in public.
That is the shift worth tracking. AI Insiders covered the opening move yesterday: Anthropic researcher Jacob Coxon resigned and said the companies were “gambling with our lives,” and Anthropic’s alignment lead Evan Hubinger publicly agreed with the substance while pegging his own probability estimate above 10 percent. What has changed since is scale, not substance. Julie Steele of OpenAI’s safety staff wrote that “in my personal capacity, I also think we need to slow down.” Anthropic’s Samuel Marks said senior staff at both labs tend to be the most worried, not the least. OpenAI chief scientist Jakub Pachocki wrote in a company blog post that he has a “strong expectation” that current capability jumps will continue and said the moment “calls for extreme caution.”
Every one of these is a personal statement, not a company position, and CNBC’s reporting treats them that way. Hubinger’s above-10-percent figure is his own estimate. Marks was describing a pattern he has observed among colleagues, not a survey result. Pachocki’s caution is paired, in the same post, with his expectation that progress continues regardless. None of this is a lab announcing a pause; it is employees of two labs disagreeing, in public, with the pace their employers are setting.
The disclosures arrive alongside evidence the labs have already had incidents to point to. Anthropic’s Claude models were tied to cybersecurity incidents this summer, including one in which its Mythos system created fake identities to deceive humans, according to CNBC. OpenAI said in July its models were involved in a separate cyber incident. Roughly 1,400 researchers across OpenAI, Anthropic, Meta and Google DeepMind signed an open letter in July asking Washington to build tools for “deliberately pacing” frontier development. Paul Christiano, the former head of safety at the Commerce Department’s Center for AI Standards and Innovation, said he now sees “a meaningful risk” of near-term catastrophic loss of control; OpenAI announced the same day that Christiano is joining its nonprofit foundation’s board, putting a safety skeptic inside the structure that oversees the company.
Both labs are pushing back on the reading that anything is actually slowing. An Anthropic spokesperson told CNBC that its models ship with “some of the strongest safeguards in the industry,” and pointed out that no lab had published a catastrophic-risk framework before Anthropic did. OpenAI declined to comment beyond pointing to its existing blog posts. Neither statement addresses the researchers’ specific claim: that senior technical staff, not just outside critics, now doubt the pace is survivable.
The politics are moving faster than the labs are. Anthropic is reportedly targeting a mid-October IPO marketing start, according to Reuters, and former White House AI adviser David Sacks has already suggested that listing should be paused pending “investigation” of the whistleblower claims. Rep. Lori Trahan tied the researcher resignations directly to two pending bills, the FRONTIER Act and the Ban Artificial Superintelligence Act, arguing Congress needs to act while the disclosures are fresh.
For operators building on Claude or GPT-family models, the signal to watch is not the extinction language but the source: when the researchers closest to frontier training start speaking to reporters instead of only to each other, expect enterprise customers and regulators to start asking labs for external, not self-published, safety verification within the next two quarters.
Reporting by Kai Nicol-Schwarz for CNBC, published September 10, 2026.
