Axios AI+

May 14, 2025
Join Axios for an event on AI, trade, and the new economic playbook on Wednesday, May 21, at 4:30pm ET in Washington, D.C., featuring conversations with Army Secretary Dan Driscoll, Sen. Jeanne Shaheen (D-N.H.), House Select China Committee Chair Rep. John Moolenaar (R-Mich.), Canadian ambassador to the U.S. Kirsten Hillman and more. RSVP here.
Today's AI+ is 1,141 words, a 4.5-minute read.
1 big thing: AI-powered hacks are a year away

Kevin Mandia, one of the most prolific cyber entrepreneurs and investors, predicts the world is only a year away from an AI agent-enabled cyberattack.
Why it matters: Mandia warned we might not even know an AI tool was the perpetrator.
- "Everybody's going to look at that, wonder how that got done, and it's probably AI behind it," he told Axios on the sidelines of the RSA Conference last month.
The big picture: AI doomsday scenarios have haunted cyber pros for decades, but the introduction of generative AI hypercharged their fears.
- Some have predicted we'll see autonomous cyber weapons that can evade security tools by 2027.
- Others predict that one day robots will be fighting robots.
Catch up quick: Mandia founded famed cybersecurity incident response company Mandiant in the early 2000s.
- Although he stepped down as CEO last year, his company has been at the front lines of major nation-state attacks and cyber criminal investigations.
- Mandia is also now a founder and general partner at Ballistic Ventures.
Zoom in: The type of attack Mandia is predicting will likely come from the cyber criminal side of the world, rather than nation-states, he said.
- Mandia added that the first iteration of any new attack style is typically "a bit sloppy" and that foreign adversaries like China are more likely to take their time before rushing to follow.
- "There is enough R&D happening right now on how to use AI [at legitimate organizations] that the criminal element is doing that R&D as well," he said.
Yes, but: Models from OpenAI, Anthropic and other popular AI companies aren't likely to be involved in the attack that Mandia is predicting.
- Those models are "pretty darn good" at blocking blatant violations of their safety parameters.
- "It's going to come from some model that's somewhere out there that's less controlled," he said.
Reality check: Chester Wisniewski, global field CISO at Sophos, told Axios that cyber criminals may already have the capabilities — but many of them don't have a real incentive to tap into them yet.
- "Fortunately today, cyber criminals are really lazy, and because we keep leaving our wallets open with large sums of cash in them, they're happy to just steal the money and move on and not do anything fancy," Wisniewski said.
Flashback: Mandia pointed to a 2001 case where the FBI arrested two Russian hackers who had programmed scripts to automate eBay purchases using stolen credit cards.
- "If they would automate all these eBay selling motions" in the early 2000s, he said, "trust me, there's already an AI agent doing this."
What to watch: Cyber defenders have been bullish on the promise of AI to help them better protect their networks and datasets.
- But their success will hinge in part on whether they can fend off the type of AI-powered attack Mandia is describing.
2. Exclusive: AI mining startup's big find
A startup using AI to guide geologic exploration believes it has found a major Australian deposit of indium, a rare metal used in solar panels, LCD screens and semiconductors.
Why it matters: Earth AI's location of what appear to be commercial concentrations shows applications for AI in the mining sector.
- In this case, it means finding indium outside of China, currently the biggest supplier, as the U.S. and other nations scramble to diversify sourcing.
Driving the news: The company, which recently raised a $20 million Series B round, told Axios exclusively that it located signs of a deposit on an outcropping roughly 310 miles northwest of Sydney.
- Some assays show up to 117 parts per million and multiple samples show over 20 ppm.
- "Ore grades for Indium typically begin at 1 ppm In, underscoring the high-grade nature of this outcrop," the announcement states.
How it works: The company has trained an AI model to predict subsurface conditions where large amounts of specific minerals and metals have likely formed.
- "A mineral deposit is like a seed in an apple. The apple is the hydrothermal system, the seed is the [mineral] deposit. It's hard to detect the seed, it's through hundreds of meters of rock, but it's easier to detect an apple," CEO Roman Teslyuk tells Axios.
- "We train an AI system to detect favorable geological conditions where deposits form," he said.
Catch up quick: It's the latest of several finds for the startup formed in 2017, including recently announced deposits of palladium, platinum and nickel.
- Its business model envisions selling rights to extraction companies once Earth AI has found deposits. It also provides services to companies for locating minerals on their holdings.
- Australia's regulatory system makes company geophysical data available, giving the firm decades of granular info that's part of a "massive training dataset," Teslyuk said.
- He calls it a major break with older, "dogmatic" exploration methods that work for shallow deposits, but now the "low-hanging fruit" has been picked.
What's next: Drilling will begin at the Kooranjie project, where the company holds rights and is already exploring other prospects there like tin and tungsten, Teslyuk said.
- Earth AI touts a proprietary drilling process that it says disturbs much less land than traditional methods.
3. VC puts big multiple on Perplexity
AI search engine startup Perplexity is nearing $120 million in annual recurring revenue and recently signed a term sheet to raise new money at a $14 billion valuation, sources tell Axios.
Why it matters: Investors are placing a high multiple on the nascent Google-killer amid a government lawsuit that threatens Alphabet's search dominance.
Driving the news: Perplexity is raising around $500 million led by Accel, as first reported by the WSJ and confirmed by Axios.
- Earlier this year, it had spoken with investors about raising $1 billion at an $18 billion valuation.
- When Perplexity closed a $500 million funding round at a $9 billion valuation six months ago, its ARR was around $50 million.
- In a break from some recent AI-related rounds, no secondary shares are expected to trade hands.
- Accel and Perplexity both declined comment.
Look ahead: Don't be surprised if Perplexity uses some of its new windfall for acquisition.
- Just about every major AI company is eyeing an add-on of some sort — whether it be for data or a direct product line — like OpenAI's $3 billion purchase of Windsurf.
If you need smart, quick intel on tech deals for your job, get Axios Pro Deals.
4. Training data
- Google is bringing its Gemini AI assistant to watches, TVs and cars. (The Verge/CNET/Gizmodo)
- Google is also testing replacing its venerable "I'm feeling lucky" home page button with an "AI mode" option. (CNBC)
- Cisco added OpenAI chief product officer Kevin Weil to its board of directors.
- AMD and Saudi Arabian AI investment company Humain will spend $10 billion on global data centers. (Bloomberg)
5. + This
Lego announced its first "Simpsons"-themed set in 10 years — a $210 brick recreation of a Krusty Burger joint that includes Sideshow Bob, Homer, Lisa and four other Springfield denizens. It's due out early next month.
Thanks to Scott Rosenberg and Megan Morrone for editing this newsletter and Matt Piper for copy editing.
Sign up for Axios AI+







