Jason Says
NVIDIA's SkillSpector is the real story today — the agent skills ecosystem is growing so fast that 26% of skills have vulnerabilities; we're building the seatbelts while the car is already doing 100mph.
NVIDIA open-sources SkillSpector, a security scanner for AI agent skills used in Claude Code, Codex CLI, and Gemini CLI. Research shows 26.1% of skills contain vulnerabilities and 5.2% show likely malicious intent — a critical safety layer as the agent skills ecosystem rapidly expands.
Kimi 2.7 launches claiming performance comparable to Claude Opus 4.8 and GPT 5.5, at 20x lower cost than SOTA models. If benchmarks confirm, it could become the best open-source model available — a massive win for budget-conscious developers and startups.
Industry insiders report Gemini 3.5 is dropping in June, with Google internally very pleased about its performance. It's predicted to be the cheapest SOTA model available — low hype now, but could reshape the competitive landscape if pricing is aggressive.
OpenAI launches three Academy courses helping employees build practical AI skills, create repeatable workflows, and deploy agents in everyday work. A clear signal OpenAI is expanding from selling models to selling AI literacy — targeting enterprise workforce transformation.
Existing memory tools like Mem0 still violate 57.5% of user preferences in coding agents. TRACE compiles corrections into runtime-enforced rules rather than passive memories — solving the core problem of agents that 'remember but don't comply.' Highly relevant for Claude Code and Cursor users.
Tool-calling LLM agents are most vulnerable to safety attacks at the very start of a session, becoming safer only after a few normal tasks — termed the 'cold-start safety gap.' New SODA benchmark measures this. Developers deploying production agents should add extra guardrails at session initialization.
Multi-agent systems lose information and waste compute by communicating through text. This paper shows heterogeneous agents with different architectures can directly share KV-cache representations, enabling dense latent communication — a foundational breakthrough that could dramatically reduce inference costs in multi-agent pipelines.
Google is suing 'Outsider Enterprise' for using AI to send 2.5 million scam texts in two weeks, defrauding hundreds of thousands. A landmark case signaling Google is weaponizing litigation against AI-powered cybercrime — an important warning shot for the entire industry on AI misuse liability.
欧洲最强开源 LLM 公司，本轮估值较 Series C 几近翻倍，显示开源模型赛道资本热度持续高涨。
史上最大 IPO，上市首日涨 19%。SpaceX 本身是 AI 基础设施（Starlink 宽带 + Grok 算力）重要载体，其 IPO 为 Anthropic、OpenAI 上市铺路。
聚焦物流、薪酬、建筑等实体经济 AI 自动化，两只基金合计 8.5 亿美元，押注 AI 落地「真实世界」而非纯软件。
Subscribe for daily AI updates + free playbook
📘 Subscribe Free