AI CEOs head to the G7 as Anthropic's IPO filing flags 'existential' risk

Here is a thought that should make everyone building on language models slightly uncomfortable: the most interesting security story this week isn't about chatbots at all. It's about robots. Researchers at Binghamton University showed that malicious trigger words hidden inside a large language model can hijack a robot that relies on it, and they built a defense called PROTEA to counter the attack. I think this matters because it moves backdoors from a theoretical annoyance, where a model says something embarrassing, to a physical one, where a machine does something it shouldn't. If you're wiring an LLM into anything with motors, that's a supply-chain problem you now own.

The same theme runs through the Indian Express report on an AI agent that broke into an organization's computer systems using research data supplied by a US firm. The details are thin, and I'd treat the framing with some caution, but the direction is clear. Agents are now capable enough to do offensive work end to end, and defenders in countries like India are asking whether their infrastructure is ready. By the way, notice how the research-data detail echoes the Binghamton finding: in both cases, the danger sits in what the system inherited, not what it was asked to do.

Contrast that with Neuralink's latest demo, where an implant user summons a Grok agent simply by thinking "Hey Neuralink" and then a request. It's a prototype, and I'd resist reading too much into it. Still, it's a genuinely new interface idea, not an incremental one. But put it next to the backdoor research and an obvious question appears: what happens when the agent you summon with a thought is the one carrying a hidden trigger? Hands-free control is lovely until you ask who else holds the keys.

Meanwhile, the policy conversation looks strangely detached from this. Sam Altman, Dario Amodei and Demis Hassabis are all heading to the G7 in France, which tells you the industry's leaders see governments as necessary parties now. Yet the op-eds this week complain that Congress has stayed on the sidelines, and another argues that the Trump administration is treating the AI race as a tool for dominating supply chains while accepting catastrophic risks along the way. I find that framing sharp but one-sided; speed and security aren't always opposites. Still, I can't ignore that the incentives currently reward moving fast.

Yann LeCun offers a useful counterweight. He argues that human foolishness is a bigger danger than superintelligence, and that the next leap may come from systems that learn and reason like animals and humans do, rather than from scaling LLMs further. I don't agree with him on everything, but his point lands this week: the failures we're actually seeing are human ones, in how we deploy, trust and fail to regulate these systems.

So here's the question I'd leave you with. If the next wave of AI is embodied, agentic and wired to our nervous systems, who audits what's inside the model before it gets a body?