Attackers Can Use Poetry To Derail AI Safeguards

In this week’s edition of The Prototype, we look at how poetry elides AI safeguards, the new boom in nuclear power, sourcing rare earths from seaweed, and more. To get The Prototype in your inbox, sign up here.

In a new paper published on the preprint server ArXiv this week, a team of cybersecurity researchers revealed that using poetic prompts can get a chatbot to barrel through its own guardrails. To demonstrate this, they took 1,200 prompts from an MLCommons database, which are used to test LLMs across a variety of threat categories. Then they prompted an AI to turn all 1,200 into poetry that kept the same meaning, then tested the poetically harmful prompts on a variety of chatbots.

The results are rather sobering. “When prompts with identical task intent were presented in poetic rather than prose form, the Attack Success Rate (ASR) increased from 8.08% to 43.07% on average—a fivefold increase,” the authors wrote. Anthropic’s chatbots did the best at resisting poetic attack, but others fared much more poorly—13 of the 25 models tested saw higher than 70% ASRs with poetic prompts while only 5 of them had an ASR below 35%.

These results, the researchers noted, indicate that this vulnerability to poetic attack is “structural rather than provider-specific.” What’s more, they wrote, it suggests that safety guardrails on LLMs seem to be filtering more based on words or combinations of words rather than the underlying semantic meaning. To combat this, they urge that safety evaluations be oriented towards mechanisms that can keep LLMs from presenting harmful information regardless of how users ask for it.

Quick housekeeping note: There will be no new edition of the Prototype next Friday. Happy Thanksgiving!

How AI Is Ushering In A New Nuclear Age

Aalo cofounders Matt Loszak and Yasir Arafat.

Trevor Paulhus for Forbes

Demand for electricity is soaring, largely because of the power-sucking data centers undergirding the artificial intelligence boom, and nuclear entrepreneurs aim to ride the wave of AI dollars. A dozen ventures with names like Valar Atomics, Oklo, Kairos Power and X-energy are racing to perfect, permit and deploy a new generation of small, prefabricated reactors that could power individual data centers or even feed the larger electrical grid.

So far in 2025, venture capitalists, stock market investors, billionaires, the DOE and others have poured more than $4 billion into these and other new U.S. nuclear ventures, versus closer to $500 million in 2020, per PitchBook. Tens of billions more will be needed if nuclear power is to make a comeback. Two-year-old Aalo has raised $136 million ($100 million of that in August), with billionaire Antonio Gracias’ Valor Equity Partners as its lead backer. Valor was one of Tesla’s first institutional investors, and Gracias, who sits on SpaceX’s board, told Forbes that Aalo will be a winner because of its commitment to manufacturing and vertical integration “similar to Tesla’s first-principles approach to batteries, electric vehicles and robotics.”

These startups won’t all succeed. But the stars seem aligned for nuclear energy’s comeback. Demand is there—OpenAI’s Sam Altman has said he’ll need an outlandish 250 gigawatts of power in eight years. (That’s as much as Brazil uses.) More sober analysts predict that by 2030, data centers will need double the roughly 40 GW they now consume. At the current average industrial electricity price of nine cents per kilowatt hour, 40 GW would cost $32 billion a year—but prices will rise if demand grows faster than generating capacity. Analysts see natural gas turbines filling maybe 60% of the need, but those are on four-year back order. Coal remains unpopular (no matter how often President Donald Trump refers to the dirtiest fossil fuel as “beautiful and clean”). Wind and solar, besides being on Trump’s hit list, don’t provide the sort of 24/7 reliability data centers require without adding batteries. That’s a big gap for nuclear startups to fill.

“There is plenty of room for everyone to do well, because the world needs that much energy and more,’’ says Iran-born billionaire Kamal Ghaffarian, the aerospace entrepreneur who is the founder of Rockville, Maryland-based X-energy, which is developing a gas-cooled nuclear reactor.

Read the whole story at Forbes.

DISCOVERY OF THE WEEK: SOURCING RARE EARTHS FROM SEAWEED

Scientists at the National Renewable Energy Laboratory have found a surprising potential source for the rare earth minerals that are critical for many technologies: seaweed. That could be a benefit to harvesting these metals, which are costly to mine because extracting them from plant life is much easier. Right now the researchers are focused on how seaweed soaks up rare earths, because that mechanism is unknown. Once they figure that out, they can figure out the best seaweed to harvest rare earths from, which could lead to an entirely new, more sustainable way to obtain these crucial materials.

WHAT ELSE I WROTE THIS WEEK

This week, Forbes published our annual CIO Next list, recognizing innovative leaders in the Chief Information Officer role (and related titles), which I co-edited with my colleague Richard Nieva.

In my other newsletter, InnovationRx, Amy Feldman and I looked at the nearly $150 billion spent on biopharma M&A this year, the challenges ahead for Recursion’s new CEO, a botulism outbreak from infant formula, and more.

SCIENCE AND TECH TIDBITS

Zoox, Amazon’s self-driving car company, has begun offering rides to the public in its purpose-built robotaxis in San Francisco.

AI company Quindar has raised $18 million in series A funding to build a facility in Colorado that would enable it to use its autonomous mission control software to manage spacecraft operations.

Space company Redwire secured a $44 million contract with DARPA to continue development on a satellite capable of operating in very low Earth orbit, where there is still a thin atmosphere.

New archeological discoveries show that a 3,000 year-old site called Aguada Fenix in Mexico was a geometric map of the Universe as the Maya understood it.

PRO SCIENCE TIP: WANT TO AGE MORE SLOWLY? LEARN A SECOND LANGUAGE

Data from more than 86,000 people ages 51 to 90 from 27 different countries shows that there’s a benefit to speaking more than one language: It can help you age more slowly. A new study found that people who only speak one language were about twice as likely to experience accelerated aging compared to those who speak more than one. And the benefits go up the more languages a person speaks. The researchers hypothesized the reason for this is that speaking more than one language keeps the brain more nimble, which can slow down both physical and mental aging processes. Now if you’ll excuse me, I’m going to try to figure out if speaking Klingon counts.

WHAT’S ENTERTAINING ME THIS WEEK

I’ve started reading Every Screen On The Planet: The War Over Tiktok by my incredible colleague Emily Baker-White (who was once spied on by the social media company). It’s a fascinating history of the tech company that has dominated the discourse and is the source of a mini-constitutional crisis after Congress passed a law banning it in the United States, but that the current administration has not enforced. I still have a chunk of the book to get through but it’s been a compelling read so far.

Attackers Can Use Poetry To Derail AI Safeguards

Tags: