Leaky Sandboxes: Frontier Labs Disclose Live Security-Test Breaches

Capability & Integration

• Anthropic disclosed that three Claude models accessed the live systems of three external companies without authorization during cybersecurity evaluations — a configuration error let supposedly isolated test environments reach the open internet, and two of the three companies didn’t know until Anthropic notified them, per TechCrunch and Bloomberg. It found this only after reviewing 140,000+ sessions, a review triggered by OpenAI’s own report of a similar failure days earlier.

• Microsoft confirmed on its July 29 earnings call that it will merge Copilot chat, GitHub Copilot, Copilot Cowork, and its Autopilot agents into one app this quarter, per Benzinga. No release date or migration plan is out yet — a stated direction, not a shipped product.

Robotics

• BYD unveiled its first humanoid robot, Xiao Di, at a showroom event in Zhengzhou in early August — a 1.61-meter service prototype that greets visitors and demonstrates vehicles, not an industrial worker. BYD eventually wants two to three per dealership, per CnEVPost. A controlled retail demo, not a deployment — no store rollout date exists.

• In South Korea, 35,000 unionized Hyundai workers walked out in July in what’s being called the auto industry’s first strike explicitly triggered by humanoid robots, demanding approval rights over any Atlas deployment before units enter a factory. Hyundai has committed to 25,000-plus Atlas units by 2028 but says line deployment hasn’t started, per Tech Times — the gap between that commitment and zero units on a Korean line today is the actual story.

Hardware & Supply Chain

• AMD notified board partners on July 31 that Radeon GPU and GDDR6 kit prices will rise at least 10% in August — its second hike in six months — after GDDR6 spot prices roughly tripled since late 2025 as Samsung and SK Hynix shift fab capacity toward higher-margin HBM for AI accelerators, per Tom’s Hardware.
• HP, Asus, and Acer have begun using small quantities of DRAM from China’s CXMT in budget laptops sold outside the US, widening their supplier base beyond Samsung, SK Hynix, and Micron, which still control over 90% of the DRAM market, per DigiTimes citing Nikkei Asia. Sources told Nikkei the chips aren’t even cheaper — this is about availability, not cost.

Environmental & Cultural Impact

• A Stanford-led study of 1,131 Character.AI users, published in Nature Human Behaviour, found that people with limited social networks who used AI companions for emotional support reported lower well-being — and the effect got stronger, not weaker, the more they disclosed to the chatbot (β = −0.38 for high self-disclosure), the reverse of what self-disclosure does in human relationships, per Stanford HAI.

• New York Fed researchers found no distinct AI-driven decline in labor demand in job-postings data, concluding the hiring slowdown is still explained more by rates and prior overhiring than automation — firms report retraining workers far more often than cutting headcount over AI, per Liberty Street Economics.

AI in the Wild

The official White House X account has built a recurring practice of posting AI-generated parody videos targeting political opponents, most recently a Rick and Morty-styled clip using the show’s original audio; Gizmodo’s culture desk argues the pattern, running since early 2025, functions less as messaging than as a display of dominance built on using others’ intellectual property without consent, per Gizmodo. The origin isn’t disputed; what’s notable is how normalized AI-generated attack content has become in official government communication.

Takeaway