What the Pope Actually Said About AI

Speaker 1 · 0:01Today on the AI Daily Brief, what the pope actually said about AI. Before that in the headlines, an update on the Mythos rollout. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. Alright, friends. Quick announcements before we dive in.

Speaker 1 · 0:25First of all, thank you to today's sponsors, KPMG, Blitzy, Robots and Pencils, and Section. To get an ad free version of the show, go to patreon.com/aidailybrief, or you can subscribe on Apple Podcasts. And if you wanna learn more about sponsoring the show, send us a note at sponsors@AIdailybrief.AI. At the end of last week, we got a quick update about project Glasswing. Now Glasswing, you might remember, was the name that Anthropic had given to the initiative through which they rolled out their new Mythos model in a very limited capacity to around 50 partners.

Speaker 1 · 0:57It was all about wrestling with the cybersecurity implications of the model, and Anthropics said the project has so far identified more than 10,000 software vulnerabilities of high or critical severity. Most partners, they wrote, have found hundreds of severe vulnerabilities in a little over a month using the model. Testing the model's accuracy on open source repos found less than a ten percent false positive rate. Wrote Anthropic, progress on software security used to be limited by how quickly we could find new vulnerabilities. Now it's limited by how quickly we can verify, disclose, and patch the large number of vulnerabilities found by AI.

Speaker 1 · 1:26Anthropic noted that industry standards demand that bugs are not disclosed for three months after identification, meaning that the comments are vague. At the same time, the report highlighted several public statements. Mozilla recently commented that they had found and fixed two seventy one vulnerabilities, which was more than 10 times the amount they found using Opus 4.6. Palo Alto Networks' most recent release featured five times as many patches as well. The report also noted that some in the finance sector are also using the model for real time fraud detection.

Speaker 1 · 1:51For one firm, Mythos was able to detect and prevent a $1,500,000 wire transfer to a threat actor who compromised a customer's account. Now one of the big criticisms of the Mythos rollout has been its exclusivity, with the model provided to only a few dozen top US companies. And in addition to the podcasters and power users of X who have been frustrated that they can't play with the new model, the frustration apparently extends to global policymakers as well. Earlier this month, European officials called on Anthropic to make mythos available to their financial sector, with similar calls having come from The UK, Canada, and governments across Asia. The Australian government held high level talks with Anthropic executives last week in an attempt to gain access.

Speaker 1 · 2:28Anthropic, for their part, says that they are currently working with critical partners, including The US and allied governments, to expand Project Glasswing and said that while they look forward to releasing a mythos class model to the public, that timeline is unclear. At present, they wrote: No company, including Anthropic, has developed safeguards strong enough to prevent such models from being misused and potentially causing severe harm. This is why we have yet to release Mythos class models to the public. Testing Catalog did spot a model called Claude Mythos one Preview being prepared for release in Claude Code and Claude Security. Their assumption is that this won't be the public release but rather broader access for trusted partners to use the model in Anthropix's harness.

Speaker 1 · 3:03So where this all stands: First, Mythos has seemingly changed the nature and speed of cybersecurity work. While these partners are reporting a 10x increase in the number of vulnerabilities detected, the bottleneck now moves to actually designing and deploying bug fixes. In their update, they wrote: Several maintainers have told us that they are currently severely capacity constrained, and some have even asked us to slow down our rate of disclosures because they need more time to design patches. The bottleneck, they continued, in fixing bugs like these is the human capacity to triage, report, and design and deploy patches for them. Box CEO Aaron Levy suggested then that the impact of Mythos is ultimately going to be a microcosm for the deployment of AI tools more generally.

Speaker 1 · 3:39He posted: This is precisely an example of why engineers don't go away, ever. We've made it far easier to create and find security issues, which means the new bottleneck is in our ability to actually review, respond to, and fix the issues. Far from AI magically solving all of this, there is still major triage work and human judgment required to do the follow on work to actually protect systems. As a result, we're about to enter a security engineer boom: Javon's paradox all over again. Now, speaking of Anthropic's partners in the government, the White House has approved a secret $9,000,000,000 budget request for intelligence agencies to build their own inference cluster.

Speaker 1 · 4:11The New York Times reports that the CIA and NSA made the budget request because they're unable to run the latest AI models on their current classified systems. The funding will go towards the purchase of NVIDIA Blackwell chips and supporting infrastructure, with the report noting government concerns about a shortage of Blackwell compute and the need to run mythos class models in settings. The New York Times suggested that the Pentagon and intelligence services have fallen behind in the AI build out after failing to allocate enough funding in recent years. After putting the issue to the White House for comment, The New York Times received what they characterized as a sharply worded statement. A White House spokesperson said: sensitive national security deliberations are conducted with the seriousness they demand, not leaked to reporters and repackaged through selectively sourced, unverified claims designed to drive headlines rather than truth.

Speaker 1 · 4:51The fact is The United States is leading the world in technology and is well prepared to deal with a variety of issues that may arise. The reporting also discussed progress in contracting between intelligence agencies and frontier AI companies. When OpenAI signed their agreement with the Pentagon in March, that contract specifically excluded the NSA. Sources said that White House Chief of Staff Susie Wiles has approved the use of anthropic models at the NSA despite the Pentagon designating Anthropic as a supply chain risk. They added that Anthropic and the government are finalizing a contract to allow the NSA to use Anthropic models in classified settings, which could help end the dispute with the Pentagon.

Speaker 1 · 5:22Intelligence officials said they hope the contract with Anthropic will also pave the way for an agreement with OpenAI that includes spy agencies. Interestingly, it seems the intelligence agencies are more flexible on guardrails than the Pentagon. The dispute with Anthropic kicked off over a limitation on using AI for autonomous weaponry and domestic surveillance. The Pentagon argued these limits could jeopardize active military operations, demanding that the contracts instead allow AI to be applied for any lawful use. The NSA and CIA, meanwhile, are technically prohibited from gathering intelligence within US borders, so are reportedly more comfortable excluding this language and agreeing to outright bans on using AI for domestic surveillance.

Speaker 1 · 5:55Now our next story is, on the face of it, just about price and token access, but clearly also has a geopolitical bent as well. DeepSeek has made their deep discount permanent as they close in on a major funding round. DeepSeek's v4 model launched in April with a 75% discount on tokens. That made the model one seventh the cost of Opus 4.6 and a quarter the cost of GPT 5.4, which were the latest models from each of the major labs at the time. Since then, Anthropic and OpenAI have each raised prices on their updated models.