# AI Safety Artifacts Open Up, NVIDIA Adds Agent Security, and Frontier Economics Sharpen *By AI News Digest • May 31, 2026* The day’s strongest signals were around the layers surrounding foundation models: safety assets moving into the open, new security tooling for agents, and a sharper debate over whether durable advantage comes from frontier scale or from the harnesses wrapped around models. ## What stood out Today’s clearest pattern sat around the stack *around* models rather than inside a single flagship model launch: more openness around safety work, more concrete security tooling for agents, and a sharper debate over whether durable advantage comes from frontier scale or from the systems built around models [^1][^2][^3][^4]. ### AI Safety Institute is reportedly putting eval assets in the open According to Hugging Face CEO Clement Delangue, the AI Safety Institute is releasing its evals, datasets, and models openly on [Hugging Face](http://huggingface.co/ai-safety-institute), so researchers can scrutinize, reproduce, and build on them [^1]. *Why it matters:* For professionals trying to separate safety claims from marketing, openly accessible artifacts make outside inspection and reuse much easier [^1]. ### NVIDIA launched a security scanner for AI agent skills NVIDIA launched [SkillSpector](https://github.com/nvidia/skillspector), a security scanner for AI agent skills [^2]. The tool is described as “Semgrep + antivirus” for agent skills and includes 64 checks across 16 categories, covering prompt injection, credential theft, supply-chain vulnerabilities, AST and taint-flow analysis, MCP security checks, optional LLM evaluation, and SARIF output for CI/CD [^2]. *Why it matters:* This is a useful signal that agent deployment is increasingly being treated like an application security and software supply-chain problem, not just a model-quality problem [^2]. ### xAI is pushing Grok Build beyond an early CLI A rollout summary amplified by Elon Musk described Grok Build v0.2.11 as moving quickly from an early CLI into a more serious agentic coding environment [^5][^6]. The update list includes integrated X and web search, export and agent commands, a read-file viewer, Always-approve mode, broader platform support, shared subagent backend services, improved context compaction, 30fps terminal video, multi-image support, and faster model switching [^5]. *Why it matters:* The release list suggests xAI is investing in persistent, tool-using developer workflows rather than a simple terminal chatbot [^5]. ### The frontier-vs-open debate is getting more economic Martin Casado argued that frontier labs are focusing on “autocatalytic” processes in which models help improve models—for example through GPU kernel creation and data cleaning—and that this should improve economies of scale [^7]. He also argued there is pricing power at the frontier, while open models face three structural challenges: pre-training is not saturated, current-generation training runs cost $2-4B, and distillation is getting harder as access to the strongest models tightens [^8][^9]. Nathan Lambert framed the same split more simply: closed models may remain slightly smarter, while open models may remain cheaper [^10]. Casado added that lagging by a few months may not matter if most value accrues to whoever stays on the frontier, and he estimated the largest frontier training runs at roughly 100,000 GPUs for six months [^3][^11]. *Why it matters:* The argument here is moving away from ideology about “open vs. closed” and toward a harder commercial question: how much value buyers place on marginal intelligence gains versus lower cost [^10][^3]. ### Harnesses are becoming part of the model story Lambert said harnesses can make models “far more independent and thorough,” pointing to a gap between weak search behavior in the Claude app and stronger task execution in Claude Code, such as pulling exact figures from papers into slide decks [^4]. In a separate interview, Gary Marcus argued that neuro-symbolic systems are “winning in practice,” because symbolic tools and deterministic code wrapped around LLMs—such as regex, loops, and Python—are doing work that pure scaling alone did not solve [^12]. > “Given that Claude seems so lazy in chat ... it seems pretty telling about how a harness can make a model far more independent and thorough.” [^4] *Why it matters:* Different camps in the AI debate are emphasizing the same operational lesson: orchestration, tools, and surrounding code can materially change what a model can do in practice [^4][^12]. --- ### Sources [^1]: [𝕏 post by @ClementDelangue](https://x.com/ClementDelangue/status/2060749008641970465) [^2]: [𝕏 post by @bibryam](https://x.com/bibryam/status/2060940955084054634) [^3]: [𝕏 post by @martin_casado](https://x.com/martin_casado/status/2060858241282723855) [^4]: [𝕏 post by @natolambert](https://x.com/natolambert/status/2060824108502565259) [^5]: [𝕏 post by @XFreeze](https://x.com/XFreeze/status/2060604454970290273) [^6]: [𝕏 post by @elonmusk](https://x.com/elonmusk/status/2060605320729186601) [^7]: [𝕏 post by @martin_casado](https://x.com/martin_casado/status/2060829314472665521) [^8]: [𝕏 post by @martin_casado](https://x.com/martin_casado/status/2060813569642602653) [^9]: [𝕏 post by @martin_casado](https://x.com/martin_casado/status/2060813284492955806) [^10]: [𝕏 post by @natolambert](https://x.com/natolambert/status/2060838705569620413) [^11]: [𝕏 post by @martin_casado](https://x.com/martin_casado/status/2060824838374301763) [^12]: [Top AI Expert: AI Won’t Replace Your Job - Here’s the Truth](https://www.youtube.com/watch?v=ZOqGaiY4ShQ)