The Right Way to Deal With AI Data Centers

**Nathaniel Whittemore** (0:00)
Today on the AI Daily Brief, the right way to deal with AI data centers. Before that in the headlines, updates on mythos, SpaceX's NeoCloud and much, much more. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.
All right, friends, quick announcements before we dive in. First of all, thank you to today's sponsors, KPMG, Robots and Pencils, Mission Cloud, and OutSystems. To get an ad-free version of the show, go to patreon.com/aidailybrief, or you can subscribe on Apple Podcasts. And to learn more about sponsoring the show, send us a note at sponsors at aidailybrief.ai.
Last note, I've been sharing the new training.bsuper.ai. This is the updated home of some of the programs that we've been experimenting with this year, including the Executive Catch-Up Program, as well as the Inheritor to EnterpriseClaw, which is the Enterprise Grade Version Executive Agent Leadership Program. That is a six-week sprint that will have you not only building agents, but also building the organizational infrastructure, policies, etc. around them that you need to actually use agents. The next cohort for that launches on June 29th. Sign up now. Again, you can find all the links that are relevant at training.bsuper.ai.
We kick off today with an update in a story that we were covering yesterday, which is around the NSA dimension of the mythos-fable ban. Now, the TLDR of what I was discussing in that initial coverage was that there seemed to me to be a mismatch between what was actually said in the way that it was being interpreted. In short, people sitting around waiting for Fable to return were looking for a reason that made a little bit more sense than there just being this jailbreak that Amazon had reported. And when they went back and found a story that seemed to indicate that the NSA had been attacked by mythos, it seemed to some to make more sense. Now, commentators with an understanding of NSA operations pointed out that what the director was telling Senator Mark Warner was likely a discussion of a controlled red team exercise rather than some live cyber attack. Shoshank Joshi, the Economist reporter whose article went viral, has now provided an update writing, A US official tells me that Senator Warner misunderstood the NSA Director General Rudd in this case. Rudd did use the hours not weeks wording, but the use of Mythos in this context was, as widely assumed, part of a red teaming effort, i.e. testing the security of internal networks. The official also told me that the agency's red teams no longer have access to Mythos because their authority for accessing it was under Project Glasswing. Joshi pointed to commentary from a cybersecurity account called IrisC2, suggesting that they had the correct read on what had actually happened. That account noted that, in the case of almost every exercise, the red team begins with some kind of initial access to the air-gapped classified network. Essentially, classified NSA systems aren't generally accessible from the outside, and gaining that initial access is one of the most difficult parts of any attack. IrisC2 commented that Mythos doesn't suddenly mean that any random person can suddenly hack the NSA. However, they did note that if an attacker manages to gain access, Mythos makes the design and execution of an exploit much faster, reducing the available time to detect and contain an intruder. So the point if we're looking for takeaways is not that Mythos has broken into the NSA, but that it does raise the stakes for cybersecurity by making attackers far more efficient. Now, the bigger detail for many was the throwaway line about the NSA not having access to Mythos anymore. Yet another reason to want this situation to be resolved as soon as humanly possible. Now, staying on the cybersecurity train, OpenAI has updated their cybersecurity model alongside a big new cyber initiative. On Monday, OpenAI announced a significant update to their Daybreak security initiative. Daybreak was first launched in May as an answer to Anthropix Project Glasswing. As part of the initial rollout, OpenAI made a preview version of GPT-55 cyber available to trusted partners. As a point of differentiation to Glasswing, OpenAI invited smaller organizations to apply for access. OpenAI wrote, Frontier defensive capabilities should not be concentrated in the hands of a few. As AI changes the pace of vulnerability discovery, defenders everywhere need democratized access to these models to find, fix and protect their infrastructure before attackers can identify and abuse these flaws. With the expansion, OpenAI has now launched the full version of GPT-55 cyber, their first model that's been fine-tuned for cybersecurity work. Like Mythos, the model has reduced guardrails to ensure professionals can use it freely in their work, and OpenAI has updated the Codex Security Plugin to make their harness more performant for cybersecurity tasks. OpenAI is claiming that with the new updates, their new cyber model has overtaken Mythos on the relevant benchmark CyberGym. In addition, OpenAI is launching a new initiative called Patch the Planet in partnership with security research firm Trail of Bits. Reporting on their initial testing of GPT-55 cyber, Trail of Bits wrote that they'd found hundreds of bugs in open-source libraries. They've deployed 37 patches so far and have many more in the pipeline. Over 30 open-source projects have already joined Patch the Planet with the goal of securing the critical software that underpins the digital world. Trail of Bits noted that the introduction of strong AI security models has fundamentally changed the nature of the work. They wrote, If it wasn't already clear from the last several months of security news, this week makes one thing clear. The expensive part of security work has moved. The advantage is no longer in finding bugs, but everything after. Confirming a finding, getting its severity right, writing a patch a maintainer will accept, and coordinating a disclosure. That is the work that floods of AI-generated reports threaten to bury. The release comes as the Five Eyes Intelligence Agencies, which is an intelligence alliance between Australia, Canada, New Zealand, the UK, and the United States, issued a rare public alert for AI-driven cyber risk. In a bulletin published on Monday, the UK National Cyber Security Center wrote, The evolving landscape of AI is rapidly transforming cyber risk and we must act swiftly to remain ahead. The bulletin called on the business community to assess the changing risk, prioritize cybersecurity practices, and stay actively engaged as threats and guidance evolve. The agencies warned that the rapid pace of frontier AI development means cyber risk assumptions can become outdated in months, not years. They also heavily encouraged the integration of AI tools into cybersecurity operations, and warned that cyber risk can no longer be treated as a purely technical issue. This is a core business risk and leadership responsibility. Now, nothing in the bulletin will be surprising to anyone who has been paying attention, but it's pretty clear that the intention of the bulletin was to light a fire under the enterprises who perhaps are not spending as much time listening to the AI Daily Brief as they should.