☕️ Google Search breaks on the word 'disregard'

Hi there, this is your daily ☕️ Techpresso.

In today's Techpresso:

🔍 Google Search breaks on the word 'disregard'

🚀 SpaceX Starship V3's first test flight was largely successful

🐛 Claude Mythos finds 10,000 critical bugs

⚖️ Google appeals search monopoly ruling, says it won business ‘fair and square’

💾 Memory prices tipped to fall as China floods the market

Plus: 🎁 10 other news you might like, 🧰 5 tools, and 📚 3 papers.

🔍 Google Search breaks on the word 'disregard' LINK

Google Search now returns a nearly empty page when users type the word "disregard," with a single unhelpful AI reply hiding the Merriam-Webster link far below a large block of blank space.

The bug surfaced after Google rolled out a new Search experience earlier this week that puts AI summaries at the top and pushes the traditional ten blue links much further down the page.

By comparison, the same search on Bing, which has taken a less aggressive approach to AI summaries, still surfaces some useful information about the word for people looking it up.

🚀 SpaceX Starship V3's first test flight was largely successful LINK

SpaceX flew its upgraded Starship V3 for the first time on Friday night, and the test flight hit most of its goals despite a couple of engine failures on both the booster and the upper-stage Ship.

The Super Heavy booster ignited all 33 Raptor 3 engines and lifted off from Starbase, Texas at 6:30PM Eastern, but lost one engine during ascent and could only do a partial boostback burn before crashing into the Gulf.

Ship reached its planned trajectory after losing one of six Raptor 3 engines, deployed 20 Starlink simulators and two modified satellites that photographed it in space, then splashed down and exploded in the Indian Ocean.

🐛 Claude Mythos finds 10,000 critical bugs LINK

Anthropic's unreleased Claude Mythos Preview model has helped partners in its Project Glasswing cybersecurity initiative find over 10,000 vulnerabilities in software in just one month, with most partners each spotting hundreds of critical- or high-severity bugs.

Bug-finding rates among partners jumped more than tenfold, with Cloudflare uncovering 2,000 bugs including 400 high or critical ones, and Mozilla fixing 271 vulnerabilities in Firefox, ten times more than with an older Claude model.

Anthropic ran Mythos Preview on 1,000 open-source projects and found 6,202 high- and critical-severity vulnerabilities out of 23,019, but won't release the model publicly until stronger safeguards exist to prevent misuse.

⚖️ Google appeals search monopoly ruling, says it won business ‘fair and square’ LINK

Google has filed its appeal against the federal ruling that labeled it an illegal search monopolist, telling the court the decision "crashed" through legal guardrails and that it "prevailed in the marketplace fair and square."

The appeal targets both Judge Amit Mehta's August 2024 monopolization decision and the September 2025 remedies order, which forces Google to share search data with rivals, including generative AI companies it says didn't exist during the relevant period.

The US and a coalition of states are also appealing, arguing Mehta didn't go far enough by declining their biggest ask, a forced sale of Chrome, and the DC federal appeals court will now decide next steps.

💾 Memory prices tipped to fall as China floods the market LINK

Memory and storage prices could soon drop because Chinese chipmakers are pumping out lots of their own RAM and storage chips, and some major PC parts brands have already started putting these cheaper Chinese chips into upcoming products.

Corsair, which usually gets its memory from US company Micron, is reportedly using chips from Chinese maker CXMT in a new 16GB Vengeance stick that still works with the standard Intel and AMD speed-boost settings gamers use.

CXMT now supplies about 7.7% of the world's RAM, sells to Chinese giants Alibaba, Tencent, and ByteDance, and saw its first-quarter sales jump 719% to $7.4 billion ahead of a planned stock market listing in Shanghai.

Other news you might like

• The memory shortage is causing a repricing of consumer electronicsLINK

• Unannounced Apple headphones revealed in FCC documentsLINK

• New battery hits 85% charge in 6 minutes without rapid degradationLINK

• How VCs and founders use inflated ‘ARR’ to crown AI startupsLINK

• Microsoft reports are exposing AI’s real cost problem: Using the tech is more expensive than paying human employeesLINK

• ClickUp cuts 22% workforce as CEO Zeb Evans pushes AI-first strategyLINK

• How SpaceX is making a quiet bet on AfricaLINK

• Cloudflare CEO Prince says builders and sellers are safe but AI is coming for the measurersLINK

🧰 Trending tools

Forsy: data infrastructure that captures and stores real AI agent workflows, helping developers debug, analyze, and improve agent behavior over time.LINK

Memdex: a Chrome extension that saves and resurfaces past AI conversations across ChatGPT, Claude, and Gemini, so you stop re-explaining context repeatedly.LINK

note.md: a local-first, privacy-focused markdown workspace for macOS that keeps your writing, research, and linked thoughts fully portable and lock-in free.LINK

RetroMac: applies retro visual filters—CRT, VHS, Game Boy, and more—to your entire screen or individual windows, with classic Dock themes included.LINK

Command A+: a customizable enterprise language model deployable across public, private, or hybrid clouds, keeping sensitive business data secure and compliant.LINK

📚 Trending papers & reports

Domain-camouflaged injection attacks fool AI safety detectors by mimicking a document's own language, dropping detection rates from 93.8% to 9.7% on Llama 3.1 8B and from 100% to 55.6% on Gemini 2.0 Flash.LINK

AI security benchmarks are fundamentally flawed, suffering from three core problems, benchmark vulnerabilities, outdated tests, and unpredictable real-world behavior, making current evaluations of AI agents unreliable and untrustworthy.LINK

AI voice assistants can be silently hijacked by hidden audio attacks, with the AudioHijack framework achieving average success rates of 79% to 96% across 13 leading models, even in contexts the attack never saw during training.LINK

Want to get the latest news differently? Find us on:

See you tomorrow for a new dose of ☕️ Techpresso!