1. 🚨 Trump Pulls His Own AI Oversight Order Hours Before Signing
May 22-24, 2026 | Ars Technica / Fortune / The Hill
President Trump abruptly cancelled an AI executive order signing ceremony after top tech CEOs — including Elon Musk, Mark Zuckerberg, and former AI czar David Sacks — lobbied him directly to kill it. The order would have created a voluntary 90-day pre-release testing regime for frontier models, which Anthropic and OpenAI had reportedly supported.
- The sequence: The EO was drafted after Claude Mythos Preview raised national security concerns. Sacks called Trump the morning of the signing. Musk and Zuckerberg followed. The event was cancelled.
- The reaction: Commerce and OSTP want light-touch regulation. The National Cyber Director wants governance now. VP Vance has voiced data privacy concerns about Mythos.
- The China dimension: Trump met Xi Jinping and agreed to launch an intergovernmental AI safety dialogue, even as the US side pulls back from domestic oversight.
- NZ angle: When the US and China both move toward AI governance (even erratically), New Zealand’s “deploy first, govern later” approach becomes more exposed.
Why it matters: The whiplash from killing Biden’s AI EO to writing his own to pulling his own — in five months — shows how fragile US AI governance is. Tech billionaires can kill regulation with a phone call. Read the full article.
2. 🚨 Claude Mythos Preview Finds 10,000+ Critical Vulnerabilities — Partners Can’t Patch Fast Enough
May 23-24, 2026 | The Decoder / IEEE Spectrum / TechTarget
Anthropic’s Claude Mythos Preview, working with ~50 partners including Cloudflare, Mozilla, Palo Alto Networks, and Microsoft, has found more than 10,000 high- or critical-severity vulnerabilities in system-critical software. The model finds bugs faster than teams can verify and patch them — creating a new category of cybersecurity pressure.
- Cloudflare: Flagged 2,000 bugs (400 high/critical) — false positive rate beat human testers
- Mozilla: Found 271 vulnerabilities in Firefox 150 — 10x what Claude Opus 4.6 caught in Firefox 148
- Palo Alto Networks: Shipped 5x the normal patch volume in its latest release
- Microsoft: Confirmed patch volumes “will continue trending larger for some time”
- UK AISI: Mythos is the first model to fully clear all simulated multi-stage cyberattack ranges
- The catch: Anthropic is keeping technical details locked down — the standard 90-day disclosure window means most findings are still under embargo
Why it matters: This is the first real-world demonstration of frontier AI as a vulnerability discovery engine operating beyond human capacity. It’s simultaneously the best cybersecurity tool ever built and a potential weapon if it leaks.
3. MIT Researcher: AI Lawsuits Could Make Courts “Grind to a Halt”
May 24, 2026 | Futurism
Self-filed lawsuits spiked from 11% to 17% by end of 2025, strongly correlated with ChatGPT adoption. MIT’s Anand Shah warns that courts “will basically have to grind to a halt” if they can’t manage the volume of AI-powered filings.
- The data: 11% → 17% self-filing rate, tracking ChatGPT adoption almost exactly
- The cost: Lawyers reporting client bills rising from thousands to tens of thousands responding to AI filings
- The paradox: Same AI that helps one person file a legitimate claim helps another file 47 frivolous motions
Why it matters: AI didn’t democratise justice — it democratised litigation. The court system wasn’t built for this volume. Read the full article.
4. AI Hallucinations in Research: Lancet Study Finds 12-Fold Spike
May 2026 | The Lancet / Fortune
A Lancet correspondence found fabricated references in biomedical papers increased 12-fold in three years. 4,000+ fake citations across ~3,000 papers. The rate went from 1 in 2,827 papers (2023) to 1 in 277 (2026). A Columbia AI researcher got caught by his own hallucination-detection tool.
- The numbers: 1 in 277 papers now has a fabricated reference (was 1 in 2,827)
- The irony: A Columbia AI researcher was caught fabricating references in his own paper about detecting AI fabrications
- The broader trend: “The Future of Truth: How AI Reshapes Reality” — a book — contained multiple fabricated quotes
Why it matters: If you can’t trust the references in peer-reviewed research, the entire citation ecosystem is compromised. The problem isn’t just that AI makes things up — it’s that the systems designed to catch mistakes aren’t keeping up.
5. AI Washing: Firms Scrambling to Rebrand as Tech-Focused
May 24, 2026 | The Guardian
UK PR executives estimate ~50% of AI-related stories they’re asked to pitch are ones they don’t want to send. AllBirds “pivoted” to AI GPUs. Standard Chartered’s CEO apologised for calling workers “lower-value human capital.” AI-powered basketball hoops and predator-protection lasers were cited as examples.
- The quote: ~50% of AI pitches are unpitchable, per UK PR executives
- The absurdity: AllBirds — the shoe company — pivoted to AI GPUs
- The damage: “AI washing” is eroding trust in genuine AI applications
Why it matters: When everything is “AI-powered,” nothing is. The hype cycle is actively harming companies doing real AI work by flooding the signal with noise.
6. EU AI Act Amendments: Deadlines Delayed, New Bans Added
May 24, 2026 | Epium
The Digital Omnibus pushes high-risk AI deadlines to December 2027 / August 2028. New prohibition on non-consensual intimate imagery and CSAM generation (effective December 2026). EU narrowed its high-risk AI list the same week the FTC broadened enforcement.
- Delayed: High-risk AI compliance pushed to late 2027/2028
- New bans: Non-consensual intimate imagery and CSAM generation prohibited from Dec 2026
- The contrast: EU narrows high-risk definitions while US FTC expands enforcement
Why it matters: The EU is simultaneously tightening some rules and loosening timelines. Companies get more breathing room on compliance, but the new bans on intimate imagery are effective sooner than expected.