Vulnerability Reports Explode as AI Bug Hunting Takes Off — Epoch AI

According to Epoch AI's latest research, 21 organizations reported roughly 1,500 high-severity and critical CVEs in June 2026 — more than 3.5 times the previous monthly record. The surge coincides with the full-scale rollout of bug hunting programs that put AI models to work.

While this demonstrates AI's potential to accelerate security research, the flood of reports is placing enormous strain on the entire vulnerability management process. New operational challenges are also a concern, including the exhaustion of remediation resources and a rise in duplicate reports. Within the security community, the key debate going forward is likely to be whether human patching processes can keep pace with AI-driven discovery.

UK AISI: Standard Benchmarks Underestimate Agent Capabilities

In a study covering seven major benchmarks, the UK AI Security Institute (AISI) concluded that standard AI evaluations systematically underestimate agents' actual capabilities. The root cause: tests impose caps on compute resources (token budgets).

Specifically, increasing the token budget 10x on software engineering tasks improved success rates by roughly 25%, and the effect was more pronounced in newer models. By AISI's estimates, the true rate of progress in frontier models is about 60% steeper than conventional measurement methods have suggested.

These findings demand a fundamental rethink of how AI capabilities are evaluated in the first place. It is significant that an official body has called out the limits of comparing models on benchmark scores alone.

Portuguese Government Releases Open-Source LLM "Amalia 9B"

The Portuguese government has released "Amalia," a homegrown large language model with 9 billion parameters. Both the SFT and DPO versions are available on Hugging Face under the Apache 2.0 license, and the paper has already been published on arXiv.

Countries and regions continue to develop LLMs tailored to their own languages and cultures, and Portugal's government-led release of a 9B-class model under Apache 2.0 is noteworthy as an open-source strategy for a smaller language community. That said, no coding benchmark scores have been published at this time.

Hong Kong Becomes the Linchpin of Asia's $2 Trillion AI Trade

According to Bloomberg's analysis of official data, Hong Kong served as the transit point for the majority of China's semiconductor imports ($239 billion) between January and May 2026. A decade ago, it accounted for roughly one-third of China's semiconductor purchases; today it handles more than half.

Hong Kong's elevated standing underscores its pivotal role in managing the flow of technology into mainland China within the AI-related trade network across Asia, now worth $2 trillion. With US-China semiconductor export controls tightening, the role Hong Kong continues to play is likely to hold the key to the global AI supply chain.

Unionization Talks at Google DeepMind Struggle to Progress

According to WIRED, negotiations toward forming a labor union at Google DeepMind are proving difficult. During a bargaining session on July 2, employees voiced frustration that management is not engaging seriously with the prospect of unionization.

While unionization efforts at AI companies partly grew out of concerns over AI safety and ethics, the negotiations themselves are not going smoothly. Whether a union actually takes shape at a major AI company could affect working conditions across the entire industry.

SK Hynix's DRAM Margins Reach 90% — Bernstein Estimate

According to Bernstein's analysis, margins in SK Hynix's DRAM business have reached roughly 90%. On Reddit's LocalLLaMA community, this has sparked debate: "If margins fell to the auto industry's typical 5%, would memory for local AI systems cost a tenth of what it does now?"

The situation has thrown into sharp relief a structure in which the memory demand boom driven by AI is delivering rare margins to a handful of suppliers. How long this lasts — or whether competition and technological change start to put pressure on it — is a critical question for the local LLM community as well.

---