OpenAI and Anthropic went head-to-head on AI jobs at the Vatican, Trump's AI security order stalled hours before signing, and India is making AI mandatory from Grade 3.
OpenAI vs. Anthropic: the sharpest public split yet on whether AI kills jobs
Sam Altman told a Commonwealth Bank of Australia audience he was "wrong" about earlier predictions of mass job losses, entry-level white-collar work hasn't been wiped out the way he expected. The same day, Anthropic co-founder Chris Olah stood before the Vatican and said the opposite: there is "a real possibility that AI will displace human labor at very large scale." Both statements were made publicly, within 24 hours of each other.
Why it matters: When the two leading AI labs can't agree on something as basic as whether their own technology threatens livelihoods, the rest of us need to stop waiting for a consensus and start preparing regardless. Both could be right, just on different timelines.
Trump's AI security order stalled: hours before he was supposed to sign it
The White House had drafted an executive order that would have required AI labs to submit powerful models to government security review (including classified testing by the NSA), before public release. On Thursday, Trump said he "didn't like certain aspects" of the order and postponed the signing with no rescheduled date. Industry and administration officials are now scrambling to figure out what happens next.
Why it matters: The absence of a rule is still a decision. US AI companies now operate without a clear government-mandated safety review process for their most powerful models, even as those models are demonstrably capable of finding and exploiting security vulnerabilities on their own.
Anthropic's Mythos model: powerful enough to find security holes on its own
The delayed executive order was partly a response to disclosures about Anthropic's internal "Mythos" model, which has been shown to autonomously identify and exploit decades-old software vulnerabilities. The NEC director compared the situation to a drug that needs to be proven safe before it goes to market. Mythos is not yet publicly released, but its capabilities are already shaping how policymakers think about AI oversight.
Why it matters: This is the clearest real-world example yet of an AI system capable of offensive cyber activity, not in theory, but in testing. The fact that it's driving executive-level policy conversations in Washington tells you where the risk bar actually sits.
India is making AI a school subject, starting from Grade 3
India's government confirmed that AI and Computational Thinking will become mandatory subjects for Classes 3 through 8 beginning with the 2026–27 academic year. The SOAR programme (Skilling for AI Readiness) is already training teachers, and Google has launched a nationwide AI educator initiative in partnership with state governments and UNICEF to build digital literacy across schools and colleges.
Why it matters: This is a systemic, top-down bet on AI literacy at a national level, not a pilot program. The next generation of Indian professionals will grow up with AI in the curriculum the way earlier generations grew up with computers in the lab.
Deepfake fraud drained $1.1 billion from US corporate accounts last year
A new SAS and ACFE study found deepfake fraud tripled from $360M in 2024 to $1.1B in 2025 in the US alone. Only 7% of anti-fraud professionals say their organisations are firmly prepared to detect AI-generated scams. Deepfake social engineering (impersonating executives on video calls, faking credentials, spoofing identity) saw the sharpest increase, affecting 77% of respondents' organisations.
Why it matters: This is not a future risk. It's happening now, in finance, in hiring, in everyday transactions. If your organisation hasn't updated its verification processes, it is already behind. Individuals are also targets: one Chicago man lost $69,000 to a scammer using a fake AI-generated US Marshals badge.
Fujitsu goes all-in on both Claude and ChatGPT for Japan's enterprise AI push
Japanese tech giant Fujitsu announced strategic partnerships with both Anthropic and OpenAI on May 27, integrating Claude and OpenAI's models into its enterprise AI portfolio for Japanese corporations. The deals are aimed squarely at large organisations managing critical infrastructure (banks, hospitals, utilities), that need AI reliability and safety guarantees, not just raw capability.
Why it matters: Enterprise AI adoption in Japan has been slower than in the US and India, but this is a major institutional signal. When mission-critical systems go AI-first, every sector connected to those systems eventually follows.
Who actually decides how safe a frontier AI model needs to be before release?
The collapsed executive order would have given the NSA a formal role in testing frontier AI models before they reach the public. Without it, no government body has that authority. The EU's AI Act has risk tiers but doesn't cover offensive capabilities like what Mythos demonstrated. Voluntary industry commitments (watermarking, shared safety testing) exist but aren't binding. Watch whether the White House returns with a revised order, or whether something else (a breach, an incident) forces the question.
Colorado's AI Act kicks in June 30 + EU AI Act omnibus deal reached
Colorado becomes the first US state to enforce a broad AI governance law when its AI Act takes effect June 30, 2026, requiring risk assessments, bias audits, and impact documentation for high-risk AI systems. Separately, the EU reached a political agreement on May 7 on an "AI omnibus" that simplifies and clarifies implementation of the existing AI Act, a deal that will reduce compliance burden for smaller companies across Europe.