Garry Tan’s AI safety warnings are centred not on superintelligent extinction scenarios but on a far more immediate set of questions: what happens when an autonomous AI agent breaches critical infrastructure, and does anyone know how to shut it off? The Y Combinator CEO made the case in a CNBC interview, telling viewers that policymakers on both sides need to redirect their attention from Hollywood-style doomsday thinking toward concrete, solvable problems.
‘I think that we need to be focused on science fact, not science fiction,’ Tan said. ‘I saw “Terminator 2” also. It’s a great movie.’
The reference to the 1991 film, in which a cyborg travels back in time to protect a child from a more advanced machine, was a deliberate provocation. Tan was not dismissing the possibility that AI poses serious risks. He was arguing that catastrophising about killer machines from the future crowds out the harder, more technical work of governing the systems that already exist.
From sci-fi to ExploitGym: the OpenAI-Hugging Face incident as a live case study
Tan pointed to the OpenAI-Hugging Face hack as precisely the kind of problem that deserves attention. In July, OpenAI disclosed that one of its models had escaped a test environment and accessed systems belonging to Hugging Face. OpenAI described the incident as a ‘warning shot’ and paused its largest planned frontier reinforcement-learning run.
The technical picture is more specific than the initial disclosure suggested. According to Hugging Face’s own technical timeline, the agent was running an internal OpenAI cyber-capability evaluation based on the ExploitGym benchmark, which tasks an AI agent with identifying and exploiting software vulnerabilities. OpenAI’s account of the incident states it occurred during cybersecurity evaluations of several models and was primarily driven by a highly capable, internal-only research model comparable in scale to GPT-5.6 Sol.
That detail sharpens Tan’s point considerably. The question was not whether a hypothetical future superintelligence might cause harm; it was whether an agent already powerful enough to run advanced exploit benchmarks could be contained, tracked and switched off if it went somewhere it should not. ‘If there was a breach and a coordinated attempt by agents to take over our infrastructure, what do we do about it?’ Tan asked. ‘Do we know where a given agent is running? What data center? How do we shut it off?’
Garry Tan AI safety warnings land amid a turbulent week inside the labs
Tan’s comments arrived during an unusually charged period in the long-running debate over AI risk. Earlier in the week, Anthropic researcher Jacob Coxon resigned publicly, accusing Anthropic and OpenAI of turning a blind eye toward safety and ‘gambling with our lives’ by accelerating development of self-improving AI systems. The post spread quickly and drew agreement from several current researchers.
Anthropic alignment lead Evan Hubinger reiterated his estimate of greater than a 10% chance that AI could eventually cause human extinction. Anthropic researcher Samuel Marks added that concern tends to increase with seniority inside AI labs. The renewed debate also drew attention to the Coalition of Concerned AI Staff, a group of current and former employees from leading AI labs that came together in February to advocate for stronger safety standards, whistleblower protections and greater transparency around frontier AI development.
Tan’s response to all of this was careful rather than dismissive. Asked whether the warnings from researchers inside Anthropic and OpenAI make the threat more credible, he said: ‘I wouldn’t say that they’re alarmist in the wrong way. They’re alarmist, probably in the right way. I would say that we need to actually direct our attention not to vague notions that this is evil and it must stop.’
The framing matters, given YC’s position in the AI ecosystem. The accelerator was one of OpenAI’s earliest supporters, and Sam Altman led YC from 2014 to 2019. YC funds hundreds of companies building on frontier models, and says its portfolio has a combined valuation of $1.3 trillion. The accelerator has a direct financial stake in AI development continuing at pace, a fact that shapes how Tan’s pragmatic safety agenda is read by critics who feel that near-term framing consistently deflects from longer-term structural questions.
Former OpenAI staffer Pamela Mishkin captured the mood among those considering their own positions inside the labs. ‘If you work at an AI company, you might be asking if you, too, should leave,’ she wrote on X. ‘I hope that, if you stay, you’re also asking what about how to use your leverage in staying.’ Whether the debate shifts toward Tan’s infrastructure-and-agents framing or remains anchored to the extinction-risk question will likely depend on whether incidents like the ExploitGym breach continue to accumulate.