# AI News Flash — Week in Review **Date:** Sunday, May 31, 2026 **Type:** weekly **Source:** https://ainewsflash.co/brief/18 **Editors:** Justin Bunnell (https://www.linkedin.com/in/justinbunnell/), Laz Manrique (https://www.linkedin.com/in/laz-m-5a218b81/) **Publisher:** AI News Flash (https://ainewsflash.co) **License:** Republish with attribution. --- ## Key takeaways - Claude Mythos' first-month Glasswing results make AI cyber-offense impossible to ignore - Nvidia's $82B print absorbed an $8B China hole and still beat — infrastructure cycle intact - Illinois SB 315 passing 110–0 shows state AI regulation has reached a coordination phase - DeepMind's AlphaProof Nexus proves AI math has crossed from benchmark play to original research - OpenAI's $4B Deployment Company and Salesforce's $1B Agentforce ARR define the enterprise agentic moment --- ## The five stories that defined the week ### Claude Mythos' first-month Glasswing results make AI cyber-offense impossible to ignore Project Glasswing's first-month results, published May 22, put numbers on what had been a theoretical claim: Anthropic's restricted Claude Mythos Preview scanned over 1,000 open-source projects and flagged 23,019 vulnerabilities, with independent firms confirming a 90.6% true-positive rate on a sample of 1,752 findings. More than 10,000 have been classified high- or critical-severity, yet under 1% have been patched — a patch-supply mismatch that is now the defining tension of the story. The UK AISI separately confirmed that Mythos and GPT-5.5 both blew past the previously measured doubling rate for autonomous cyber-task completion, and Forrester noted this week that Mythos effectively obsoletes nation-state zero-day stockpiles by making discovery cheap and systematic. The structural shift is this: Anthropic has built a model it considers too dangerous to sell, is giving it to roughly 50 vetted partners for defensive use, and has privately warned US government officials that Mythos makes large-scale cyberattacks significantly more likely this year. Watch next week whether CISA or Commerce issues any public guidance in response — Anthropic said it is working with US and allied governments to expand the program. - Source: https://www.anthropic.com/glasswing ### Nvidia's $82B print absorbed an $8B China hole and still beat — infrastructure cycle intact Nvidia's Q1 FY2027 report was the clearest test yet of whether the AI infrastructure buildout survives the US-China export wall: it does, for now. Revenue of $81.6 billion came in $2.8 billion above consensus despite an estimated $8 billion gap from the H20 China restrictions, with data center at $75 billion and Blackwell described by Jensen Huang as 'off the charts' with cloud inventory sold out. The stock's ~1% post-earnings slip continues a pattern visible since late 2025 — buy-side desks have internalized that a routine beat is already priced, and the real question each quarter is GB300 Ultra ramp timing and Q2 guidance headroom. The Oracle-OpenAI $300B Stargate deal, confirmed this week as targeting 4.5 GW of capacity by 2027, is the demand backstop behind that guidance — hyperscaler capex commitments for 2026 now exceed $700 billion combined. The China hole is structural rather than cyclical: management has excluded the region from forward guidance since H20 restrictions took effect, and any path back requires a policy reversal that neither Commerce nor the White House has signaled. Watch Q2 guidance validation as GB300 Ultra moves from sampling to production shipments. - Source: https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-first-quarter-fiscal-2026 ### Illinois SB 315 passing 110–0 shows state AI regulation has reached a coordination phase The week's regulation story is best read as a single coordinated arc rather than a stack of isolated bills. Illinois passed SB 315, the Artificial Intelligence Safety Measures Act, 110-0 in the House on May 27 — it would become the first state law mandating annual third-party safety audits of frontier AI developers and is now with Governor Pritzker. This happened on the same week New York's algorithmic-discrimination bill (S 1169) cleared committee with a private right of action, and as the Illinois Senate pushed a full eight-bill AI package to the May 31 close. Illinois sponsors explicitly cited California and New York by name, arguing the three states would cover roughly 40% of the US AI market — the clearest public articulation yet that large-state legislatures are treating themselves as a de facto federal substitute. The federal backdrop makes this sharper: a White House executive order earlier this year directed the Attorney General to set up an AI litigation task force to challenge state laws on preemption grounds, and that tension has not been resolved in court. What to watch: whether Pritzker signs SB 315 (Anthropic backed it; most big tech opposed it), and whether a Commerce Department preemption evaluation triggers a legal challenge before any of these laws take effect. - Source: https://capitolnewsillinois.com/news/senate-democrats-introduce-bills-to-regulate-artificial-intelligence/ ### DeepMind's AlphaProof Nexus proves AI math has crossed from benchmark play to original research AlphaProof Nexus solved 9 of 353 open Erdős problems and proved 44 of 492 open OEIS conjectures — published on arXiv May 21 with all proofs machine-verified in Lean 4 and posted to GitHub for anyone to check. The cost per solved problem was a few hundred dollars, and crucially every proof compiles: no peer review required, no ambiguity about correctness. This is structurally different from OpenAI's prior GPT-5.2 Pro Erdős result, which was a natural-language argument that required human mathematicians to confirm and later surfaced a precedent in the literature. DeepMind CEO Demis Hassabis moved quickly to frame the result as a research workflow advance rather than AGI, but the underlying dynamic is harder to dismiss: an agentic loop pairing Gemini 3.1 Pro with Lean's formal checker is now producing original mathematical discoveries at computation costs that make sustained research campaigns cheap. Carnegie Mellon's Jeremy Avigad wrote in March that the discipline of mathematics 'must engage with these tools before they outpace it' — this week's result is the evidence. Watch whether the team extends the system toward Millennium Prize problems, as the paper's conclusion explicitly names that as the next target. - Source: https://arxiv.org/abs/2605.22763 ### OpenAI's $4B Deployment Company and Salesforce's $1B Agentforce ARR define the enterprise agentic moment Two announcements this week, read together, define what the enterprise AI market actually looks like in mid-2026. OpenAI spun out a standalone Deployment Company backed by $4 billion from TPG, Brookfield, Advent, and Bain Capital, with Forward Deployed Engineers embedded directly inside customer organizations — a services motion that acknowledges selling API access is not sufficient to capture enterprise budgets. On the same week, Salesforce reported Agentforce ARR crossing $1 billion, growing over 200%, with 3.8 billion Agentic Work Units delivered in Q1 — the first credible public proof point that agentic AI is generating recurring software revenue at scale, not just pilot spend. The tension between these two data points: Salesforce stock is still down 33% year-to-date on concern that AI disrupts its own core software. OpenAI's Deployment Company is the mechanism by which AI-native vendors reach into the same enterprise workflows Salesforce has owned for two decades. Whether OpenAI's FDE model can scale — and whether Salesforce can defend its installed base with Agentforce before those FDEs displace it — is the enterprise AI story for the second half of 2026. - Source: https://openai.com/index/openai-launches-the-deployment-company/ --- Archive: https://ainewsflash.co/archive RSS: https://ainewsflash.co/rss.xml LLM index: https://ainewsflash.co/llms.txt