Here are six key stories from overseas AI news on August 5 (Japan time). The lineup covers the gap between open-weight models' capability and safety, mixed signals across the semiconductor market, the reality of AI security, code review hitting a wall, growing autonomy in military AI, and a field report on running local LLMs.

Open-Weight GLM-5.2 Nears the Frontier, but Safety Lags Behind

A new report from SaferAI notes that Z.ai's open-weight model GLM-5.2 is closing in on frontier-class capability while lacking key safety measures. It renews concerns that powerful open models are running ahead of the buildout of governance and guardrails. The democratization of capability is moving fast, but operational maturity still lags behind. The smoldering debates over generative AI in the open-source community also appear to be rooted in this gap between capability and governance.

AMD's Q3 Revenue Outlook Falls Short of Market Expectations

AMD announced that third-quarter revenue is expected to come in at around $13 billion, plus or minus $300 million. While that beats the analyst consensus of $12.5 billion, some Wall Street estimates had run well above $13 billion, so the guidance was widely seen as underwhelming relative to expectations for AI data center expansion. AI demand continues to drive semiconductor earnings across the board, but this is also a moment where the sheer height of those expectations is being tested.

AI Guardrails: 'Even a Script Kiddie Can Break Through'

Reports keep piling up that bypassing guardrails (safety restrictions) is startlingly easy. The Register described the current state — where clever prompting and simple techniques slip past protections — as something even a script kiddie can do. Separately, there are reports that many AI-built apps expose credentials. One tool scanned 256 AI-built apps and found that the majority leaked credentials in some form. Model-level safety matters, but the operational risks on the application side are the ones most likely to surface first.

AI Writes the Code, Humans Can't Review It All

As code generation becomes increasingly automated, the review process is struggling to keep up. In an era when AI writes code at scale, having humans scrutinize every line is no longer realistic — and the question of what to do about it is now front and center. Some voices take a positive view, arguing that AI actually reduces the harm caused by less capable engineers. Meanwhile, tools for running multiple agents at once and managing them safely (such as Watchfire and AgentRails) have started to appear, and teams on the ground are shifting toward building out monitoring and control mechanisms.

US X-62A: AI Autonomously Intercepts an Enemy Aircraft

AI autonomy has taken another step forward in the military domain. The US experimental aircraft X-62A, equipped with an infrared search-and-track pod, reportedly used an AI "brain" to autonomously intercept an enemy aircraft (a T-38). Demonstrating an autonomous intercept without pilot input suggests that AI's role in aerial combat is expanding from training support into territory closer to actual combat. That said, details on operational deployment and standards remain limited for now.

(Addendum) Full Kimi K3 Model Reported Running at 20+ tps on a GB10 Cluster

In the local LLM community, a report surfaced of someone running the full Kimi K3 model on a cluster of 16 NVIDIA GB10 units, recording an average of 20+ tps (peaking at 38 tps, with prefill inference at 750 tps). The poster says they plan to release a vLLM image and instructions. It's an individual field report rather than confirmed information, but it shows that efforts to run large-scale models in local environments are ongoing.

---

Capability leveling out, a gap between infrastructure and expectations, the reality of safety — whichever layer you looked at, the same pattern kept showing up today: AI runs ahead, and the human side scrambles to catch up.