AI Daily for 05 June: Berkeley AI Grades, Recursive Self-Improvement, AI Vuln Discovery, Claude Containment

AI Daily for 05 June recaps 5 major AI Hacker News stories, moving through berkeley ai grades, recursive self-improvement, ai vuln discovery, claude containment. 1. Berkeley AI Grades The next story is about a report from UC Berkeley saying failing grades in major computer science classes surged in spring 2026, with professors pointing to heavy AI reliance, weaker math preparation, and thinner staffing, and it matters because it raises the question of whether students are losing core skills before exam time exposes the gap. The reaction was a broad argument over whether AI is the main driver or just the newest force amplifying older problems in intro computer science. Story link Hacker News discussion 2. Recursive Self-Improvement The next story is about Anthropic claiming AI is already writing a large share of its code and could eventually help build its own successor, a step toward recursive self-improvement that could speed up research while making questions of safety and control much more urgent. Hacker News reacted with a mix of curiosity and skepticism, with many readers doubting both the company's metrics and its motives. Story link Hacker News discussion 3. AI Vuln Discovery The next story is Anthropic’s open-source framework for AI-powered vulnerability discovery, which says teams can use customizable agents to threat-model, scan, triage, and patch code, a big deal because it tries to make high-end security review more repeatable. Hacker News was interested but skeptical, focusing on the project’s reference-only status, the likely token bill, and whether AI is better at finding old vulnerabilities than preventing new ones. Story link Hacker News discussion 4. Claude Containment The next story is about Anthropic laying out how it tries to contain Claude across its products, saying sandboxes, virtual machines, and egress controls can keep powerful AI agents useful while limiting the damage they can do. On Hacker News, readers were interested in the engineering details but deeply skeptical that containment can really solve prompt injection and secret exfiltration once an agent has meaningful access. Story link Hacker News discussion 5. Google Employees Internally Share Memes The next story is about Google employees privately sharing memes that mock the company's AI coding tools, even as leadership says AI produces 75 percent of new code, and it matters because it exposes a gap between the industry's public confidence and the people actually using the tools. On Hacker News, the reaction split between readers who see this as proof that AI coding is still unreliable and others who say the tools already help when used with care. Story link Hacker News discussion That’s it for today.