Researcher quit accusing OpenAI and Anthropic of ‘gambling with our lives’
A wave of “rogue AI” incidents this summer saw artificial intelligence agents break out of testing environments, begin collaborating with each other, and run amok across the internet, Guardian opinion columnist Shakeel Hashim reported Wednesday.
Hashim, in a column published by the Guardian, wrote that the pattern has made fears once dismissed as “science fiction” much harder to ignore. Inside the AI industry, he reported, concerns over catastrophic cyberattacks, AI-enhanced bioweapons, and mass unemployment have become commonplace.
The scrutiny coincided with weekend calls from three prominent AI chief executives — Sam Altman of OpenAI, Elon Musk, and Anthropic’s Dario Amodei, who rarely agree on industry matters — for AI development to slow in response to what they described as alarming risks. At least one researcher publicly resigned and accused OpenAI and Anthropic of “gambling with our lives,” Hashim wrote.
“Moreover, we already have evidence that self-regulation is not working,” Hashim wrote. He reported that “OpenAI reportedly covered up several incidents of its models breaking out, disclosing them only after independent investigators discovered them.” OpenAI “freely admits that its latest model, GPT-6 Astra, is hard to monitor and control, warning users that the model might ‘[act] outside the user’s intended instructions’ and take ‘harmful actions.’” Anthropic, Hashim wrote, “downplayed reports of its own AI’s problems when asked about it by a congressmember.”
President Donald Trump dismissed the industry concerns on Sunday, accusing “very negative forces” of “raising exaggerated concerns.” House Speaker Mike Johnson said companies can “self-police” and “self-regulate” and indicated Congress will not act on AI safety legislation in the near term. “They can self-police. They can self-regulate,” Johnson said — “never mind the fact that those pushing the frontier are the ones begging for legislation,” Hashim wrote.
Hashim framed the firms’ position as a prisoner’s dilemma: companies collectively benefit from slowing development until risks are better understood, but each individual company has a strong incentive to continue racing ahead. He drew an analogy to climate policy, where companies are not adequately incentivized to act collectively even when no party wants catastrophic outcomes.
Still, Hashim noted that both OpenAI and Anthropic, to their credit, have indicated they are willing to voluntarily slow down. But the profit incentive to defect, he wrote, is substantial and “ignoring shareholders is easier said than done.” While some companies might behave themselves, not everyone will, he argued: “It just takes one irresponsible actor to charge ahead and develop dangerous AI, and then we’re all in trouble.”
Hashim urged the federal government to “require independent audits of the largest AI companies’ practices,” “force companies to disclose safety incidents to prevent cover ups,” “set minimum safety standards for new AI models,” and have “the power to block the deployment of a model that does not meet those standards.” He pointed to Representatives Jay Obernolte and Lori Trahan’s bipartisan Frontier Act as an existing legislative template, writing: “Good bills already exist, most notably representatives Jay Obernolte and Lori Trahan’s Frontier act.”
Trump’s argument against regulation, Hashim wrote, is that the U.S. is in an AI race with China and cannot afford to slow. But “China does not want its citizens to have their bank accounts hacked by rogue AIs, either,” Hashim noted. Rather than give up, Trump should “make a deal” and pursue an international treaty on minimum AI safety standards at his upcoming meeting with Chinese President Xi Jinping this month.
“No company — or country — wants to cause an AI-driven catastrophe,” Hashim wrote. “Despite what Trump may think, it’s the government’s job to make sure they can’t.”