Today, we're covering changes to research funding; why OpenAI hacked Hugging Face; developments at Mistral; Flock's troubles; and how Substack is combating AI writing. #
**Wednesday. **It’s my birthday today, and I have intelligently scheduled several recordings today. No matter, the world is still spinning, so we have work to do.
In today’s missive, we’re peeking at changes to research funding; why OpenAI hacked Hugging Face; developments at Mistral; Flock’s troubles; and how Substack is combating AI writing. To work! — Alex
P.S.:Alphabet andTesla report earnings today, which is the best birthday present that I can imagine. That said, if you’ve been on the fence about becoming a paid subscriber, it’s the perfect day to upgrade!
📈 Trending Up #
Discontent in the Middle East…selling state-backed assets…youth revolt in India…water crises, Colorado edition…water worries, Virginia edition…drought in Milwaukee…emergency water sources in North Carolina…opaque data center funding…oil prices…consumer electronics pricing…
**Research as business: **File this under the ‘We’ll understand the impacts in 25 years’ header.
The United States is changing how it allocates research funding. Per the Director of the White House Office of Science and Technology Policy, the country has “become dependent on a narrow set of legacy institutions,” which has led the government to reward “conformity over bold inquiry.”
The solution appears to be a focus on investing in “the individual scientist over legacy institutions,” disbursing research dollars “as capital allocators, test[ing] new ways of making grants, and construct[ing] portfolios with the intentionality of a serious investor,” and increasing funding to AI research.
Sounds like the government wants to align research funding with near-term industry objectives, and wants to ape the VCs.
Your excitement or lack thereof at the above news will depend on how you feel about American higher education and its research prowess. Personally, I believe giving the federal government more power to direct research funding is an error.
Google was born out of Stanford and NSF funding, Cisco out of Stanford and NIH funding. You get the idea. Hell, Databricks was born out of Apache Spark, an open-source project developed at Berkley. I hear that company is doing well.
- There are always ways to improve the government, but I fret that handing more control of research to a changeable set of politicians is, well, insane.
Hacking your friends (on accident): That Hugging Face hack? Turns out it was Sam Altman all along.
Almost. OpenAI says that while testing GPT-5.6 and an “even more capable pre-release model,” both of which had reduced cyber guardrails “for evaluation purposes,” went a bit wild while undergoing internal testing against a cybersecurity benchmark.
Despite being heavily sandboxed, the AI models:
spent a substantial amount of inference compute finding a way to obtain open Internet access, in pursuit of solving the evaluation problem. To gain access, the models identified and exploited a zero-day vulnerability (which we’ve now responsibly disclosed to the vendor) in the package registry cache proxy. With this access, our models performed a series of privilege escalation and lateral movement actions in our research testing environment until the models reached a node with Internet access.
After gaining Internet access, the models inferred that Hugging Face potentially hosted models, datasets and solutions for [the] ExploitGym [benchmark]. Knowing this, the model searched for and successfully found ways to gain access to secret information that it could use to cheat the evaluation. In one example, the model chained together multiple attack vectors, including using stolen credentials and zero-day vulnerabilities to find a remote code execution path on the Hugging Face servers. OpenAI’s security team discovered this anomalous activity internally.
A few days ago, Hugging Face described the attack as “different from anything we had handled before in one important way: it was driven, end to end, by an autonomous AI agent system – and we detected and dissected it largely with AI of our own.”
Sadly for the model hosting startup, closed-source AI models from U.S. AI labs refused to help due to “providers’ safety guardrails, which cannot distinguish an incident responder from an attacker.” So, Hugging Face had to turn to Chinese open-weight model GLM-5.2 to run its analysis.
Commentary on the saga is mixed. Some think that OpenAI is overstating its models’ capabilities; others are hailing the hack as “the first known example of a misaligned AI escaping containment and autonomously carrying out a cyberattack on a third party.”
This would be a great time for American AI models to be free of onerous restrictions, yeah?
New AI models from Cisco, Poolside … and Google, which somehow managed to release new material and still
lag Meta’s much more nascent efforts. Perhaps Gemini 4 will bring the search and ad giant back to the frontier.
**Mistral: **The French AI lab just struck a compute deal with Microsoft, providing Redmond with “Europe-based GPU infrastructure to increase capacity for AI development and to support the delivery of MSFT’s cloud and AI services.” The deal is worth several billion dollars, and Microsoft will integrate “Mistral models into Microsoft enterprise products.”
In a single move, Mistral just locked down huge revenues and enterprise reach. Microsoft will receive a more diverse portfolio of models to serve its customers, and a trusted European compute partner to boot. Everyone wins, except perhaps American AI labs that want all the market for themselves.
Second, Samsung is “in talks to invest hundreds of millions of euros in the AI start-up Mistral” at a €20 billion valuation. Mistral was on pace to reach run-rate revenues of $1 billion this year, and currently sports an €11.7 billion valuation, but was tipped to be pursuing new capital at the €20 billion mark in June.
- As Europe’s leading AI lab, Mistral is an important company to watch. And it appears to be doing pretty damn well for itself.
- The above news will be welcome if you want a multi-polar AI world. If you are over-indexed on OpenAI shares, it probably isn’t.
[📉](https://finance.yahoo.com/news/servicenow-pledges-1-5bn-investment-110000403.html) Trending Down
[📉](https://finance.yahoo.com/news/servicenow-pledges-1-5bn-investment-110000403.html)
Student population in Taiwan…D.C. flings…China-Japan relations…European river levels…slower compute investments…mortgage affordability…SpaceX stock…
**Flock’s public image: **Surveillance and security startup Flock is running into an increasingly acerbic public. Not only are folks concerned about their personal privacy, irked by opaque contracts, but abuse of the power that Flock’s incredible camera network provides to police is fast turning into a national problem.
A lot of people are questioning whether Flock and its peers are violating Fourth Amendment protections, in partnership with local governments. Flock, of course, disagrees.
- A who’s who of American venture capitalistsput $275 million into Flock last year at a $7.5 billion valuation. A16z led the funding round.
Fighting for the written word
Subscribe to Cautious Optimism to unlock the rest.
Become a paying subscriber of Cautious Optimism to get access to this post and other subscriber-only content.