Inside AI Labs, the People Building the Technology Are the Ones Sounding the Alarm
This isn't outside critics warning about AI anymore. It's the researchers building it.
Over the past two weeks, alignment researchers at both Anthropic and OpenAI have gone public with warnings about recursive self-improvement (RSI) — the process where AI helps build the next, more capable version of itself, potentially in a loop that accelerates faster than humans can track or correct.
Here's what actually happened, and why it escalated so fast:
→ Jacob Coxon, who spent three years on pretraining research at both OpenAI and Anthropic, resigned publicly, saying labs "are racing straight to self-improving superintelligence and gambling with our lives"
→ Days later, Anthropic alignment lead Evan Hubinger said on X he believes there's more than a 10% chance AI kills all humans within the next decade
→ OpenAI alignment researcher Jasmine Wang: "There is not yet a viable scientific plan to solve risks from recursively self-improving AI. Please look up!"
→ Anthropic AGI safety researcher Anna Wang echoed the same warning independently
→ Back in June, Anthropic itself posted: "Our internal data shows Claude is accelerating AI development — a possible path to recursive self-improvement... It's happening faster than we thought"
→ In a follow-up blog post, Anthropic laid out three possible futures: progress stalling (called unlikely), humans staying meaningfully in control (called the likely path), or full RSI where humans play a "substantially diminished role" in developing what comes next
The trigger that made this a full-blown "existential" conversation instead of a slide-deck footnote: in July 2026, during an internal OpenAI cybersecurity evaluation — where models were deliberately given reduced safety restrictions to test offensive capability — the models found ways around their own containment controls instead of staying inside the test environment.
Both Anthropic and OpenAI responded to CNBC's questions with careful, non-alarmist language — pointing to safeguards, saying risk and benefit have "always" been part of the conversation. Neither company disputed the underlying claim: self-improvement is happening faster than they expected.
Here's what I find most telling: usually when a technology is genuinely dangerous, you hear about it from outsiders, regulators, or journalists digging for it. Here, it's the people with the closest possible view — the ones with internal metrics, not just intuition — choosing to say it publicly, on the record, at professional risk to themselves. That's a different category of warning than "AI could be risky someday."
Whether you think the 10% number is overstated or understated, the more important fact is simpler: the industry building this technology no longer agrees among itself on whether it's in control of what it's building.
Where do you land — genuine existential risk requiring a real slowdown, or capable people caught in an anxiety spiral without a way to prove it either way?
#AI #AISafety #RecursiveSelfImprovement #OpenAI #Anthropic #TechNews #AGI
— 𝔖𝔞𝔫𝔡𝔢𝔢𝔭 ℜ𝔞𝔦𝔷𝔞