Daily Briefing — April 8, 2026

Anthropic just announced Project Glasswing, an AI model that has apparently found security vulnerabilities in every major operating system and web browser. That is not a small claim. The model, called Claude Mythos Preview, was not specifically built for cybersecurity but turns out to be exceptionally good at it anyway, thanks to its reasoning and agentic coding capabilities. Anthropic is keeping it locked to a tight circle of "defensive security" partners including Nvidia, Google, AWS, Apple, and Microsoft, precisely because the same tool that finds vulnerabilities could absolutely be weaponised to exploit them.

Here is the part that deserves your attention. This is not a niche security product for a niche security team. This is a general purpose model that stumbled into being genuinely dangerous in the best possible way. That tells you something important about where AI capability development is heading: the gap between "useful assistant" and "nation state level threat analyst" is getting uncomfortably thin.

The restricted rollout is a deliberate choice, and a revealing one. Anthropic is essentially admitting that releasing this publicly would be like handing a lockpick set to everyone on the street and hoping most of them are locksmiths. The fact that major tech companies are lining up as partners suggests this is not a PR stunt. Organisations are genuinely scrambling to get ahead of AI powered attacks before the offensive side of this equation becomes democratised.

Google's AI Overviews is wrong about one in every ten questions it answers. That sounds almost acceptable until you remember that Google processes something in the range of 8.5 billion searches per day, and AI Overviews shows up on a massive chunk of those. Do the math and you are looking at tens of millions of incorrect answers going out every single day to people who are treating that top result like gospel truth.

The New York Times ran this analysis with help from an AI startup called Oumi, using a benchmark called SimpleQA, which is basically a standardised test of over 4,000 factual questions with verifiable answers. When Gemini 2.5 was running the show, accuracy sat at 85 percent. After the Gemini 3 update, it nudged up to 91 percent. Progress, sure. But "improved from terrible to merely bad" is not the victory lap Google's marketing team would write home about.

The deeper problem here is not really the number. It is the confidence. AI Overviews does not say "I think" or "I am not sure." It just tells you things. It sits at the top of the page, above every other result, in a format that signals authority. And most people are not cross-referencing. They read it, they move on, and a bad answer is now rattling around in their head as fact.

SpaceX is gearing up for what could be the largest IPO in history, targeting a $2 trillion valuation and aiming to raise $75 billion when it finally goes public. The company has scheduled a June roadshow starting with analyst briefings on the 7th, followed by a retail investor event on the 11th for 1,500 regular people. That last part is the unusual bit.

Up to 30% of shares could go to non-institutional investors, which is a genuinely rare move for a private tech company of this scale. Normally at this level, the big institutional players get first access and everyone else reads about it in the news the next morning. SpaceX is deliberately flipping that script, leaning hard on Musk's retail fan base to fuel demand the same way it fuels his other ventures.

Patrick Ciriello has a master's degree, decades of experience in IT and software systems, and a track record across banking, pharma, and universities. He also spent the better part of a year getting ghosted by deli counters. That sentence should make you stop for a second. This is not a story about someone who coasted or got complacent. This is someone who survived the dotcom crash, 2008, and Covid, rebuilt each time, and still ended up on the wrong side of a job market that has quietly decided experience has an expiration date.

What the Guardian is documenting here is something a lot of people in their 40s, 50s, and 60s are living right now but not saying out loud because there is still a stigma attached to being skilled and unemployed at the same time. AI training has become the thing people turn to when everything else stops working, a kind of career life raft that the industry has handed workers without being totally honest about whether it actually floats.

The deeper implication is uncomfortable. If someone with Ciriello's background is sending out hundreds of applications and hearing nothing back, the problem is not a skills gap. It is something structural, and layering AI certifications on top of it may not fix the underlying dynamic.