Show notes — Episode 4
The week of 28 June 2026. A source under every claim, so you can dig in and check the work. Points ordered by long-term impact.
1. The US government now decides, company by company, who can use frontier AI
Two weeks ago the Commerce Department's Bureau of Industry and Security barred foreign nationals, inside or outside the US and including Anthropic's own non-citizen staff, from using Anthropic's Mythos 5 and Fable 5 models, citing a jailbreak that could bypass safeguards to reach the models' cybersecurity capabilities. Because the platform cannot verify nationality in real time, Anthropic took both models offline for everyone. On Friday 26 June the Commerce Department partially reversed the order: according to a letter from the Commerce Secretary that NPR has seen, the government will now let a vetted list of American companies use Mythos 5, including those companies' foreign employees; Fable 5 remains restricted. The structural change is who controls the list: the government, not the company. The same week, OpenAI said it had agreed to let the administration vet which companies get access to its newest model, calling that "the strongest path to broader availability" while saying such vetting "should not be the norm long term." So within two weeks the US has moved from a hands-off posture to per-company pre-clearance of access to the most capable models, with one lab compelled by an export order and a second adopting the same gatekeeping voluntarily. NPR notes Mythos has had preview access at companies including Google, Microsoft and Cisco; Cisco's head of security said it scanned 1.8 billion lines of code in eight weeks, work he estimated would otherwise take eight-plus years. The counter-current: on 13 June the Chinese lab Z.ai released GLM 5.2, an open-weight (MIT-licensed) model that independent coding benchmarks place at roughly the level of the top closed frontier models, a few points behind Claude Opus 4.8 (e.g. SWE-bench Pro 62.1 vs 69.2; near-tie on FrontierSWE). The structural read is not the price or the margin but the direction: open-weight capability is reaching the frontier just as access to the closed frontier is being restricted by export law.
Sources: NPR (primary: 26 June partial reversal, letter seen by NPR, vetted US-company list, OpenAI agreeing to vetting, Cisco's 1.8bn lines), CNBC (Mythos 5 released to a group of companies and federal agencies), CSIS (12 June BIS order, EAR authorities, first use of this mechanism on model access), Tech Policy Press (export-control precedent via API-delivered models), VentureBeat and The AI Rankings (GLM 5.2: 13 June, open weights, coding benchmarks vs Opus 4.8).
2. The EU is thinning and delaying its AI Act before the main rules take effect
The EU agreed the first amendments to the AI Act since it passed in 2024, packaged as the "Digital Omnibus." A provisional agreement among the Council, Parliament and Commission (reached 7 May 2026, endorsed by Parliament 16 June) postpones the high-risk obligations: stand-alone Annex III systems (AI used in areas such as hiring, education and critical infrastructure) move to 2 December 2027, and AI embedded in regulated products under Annex I to 2 August 2028. The original applicability date was 2 August 2026. What still comes due on 2 August 2026 is the Article 50 transparency duty (telling people when they are interacting with AI or AI-generated content), and the package adds a new Article 5 prohibition on AI tools that generate non-consensual intimate imagery ("nudifiers") and child sexual abuse material. Analysts describe the deferral as an acknowledgement that the regulatory infrastructure to make the high-risk rules operable did not materialise on schedule. The result: the world's first comprehensive AI law is being softened and delayed at the same moment the US is assembling a control regime out of export law.
Sources: Gibson Dunn (Digital Omnibus deadlines: Annex III to 2 Dec 2027, Annex I to 2 Aug 2028), ComplianceHub (Article 50 transparency remains live for 2 Aug 2026), European Commission (AI Act framework and timeline), Covington / Global Policy Watch (new Article 5 prohibition; first amendments since 2024).
3. Colombia elected a Trump-backed far-right outsider
Abelardo de la Espriella, a 47-year-old lawyer who has never held elected office and reached the ballot through citizen signatures rather than a major party, won Colombia's presidential runoff with about 49.7% to left-wing senator Iván Cepeda's 48.7%. The National Electoral Council confirmed the result on 24 June 2026. He campaigned on a culture-war platform, casting himself as a defender of the "traditional family" and opposing abortion, same-sex adoption and "gender ideology," and he was endorsed by US President Donald Trump. The result returns the presidency of a major Latin American country to the right after Gustavo Petro's left-wing government and adds another Trump-aligned leader in the hemisphere for the next four-year term.
Sources: Al Jazeera (49.7% vs 48.7%, citizen-signature ballot, culture-war platform), CNN (Trump endorsement, razor-tight runoff), Wikipedia: 2026 Colombian presidential election (National Electoral Council confirmation 24 June).
4. Keir Starmer resigned as UK Prime Minister
Keir Starmer resigned as Prime Minister and Labour leader on 22 June 2026, under pressure from his own party. The frontrunner to succeed him is Andy Burnham, mayor of Greater Manchester, who won a by-election in the Makerfield seat on 18 June to return to Parliament shortly before Starmer quit. Labour's National Executive Committee set nominations to open on 9 July, with the contest to conclude before the summer recess; the Eurasia Group projects Burnham could take office around 18–19 July. If he wins, he would be the UK's seventh prime minister in roughly a decade. The change of leader comes less than two years after Labour won a large parliamentary majority.
Sources: NPR (resignation 22 June, seventh PM in a decade), CNN (Burnham frontrunner, Makerfield by-election 18 June), Al Jazeera (nominations 9 July, timeline to recess).
5. The US and Iran agreed a 60-day roadmap to a final deal, with IAEA inspectors returning
The US and Iran concluded their first round of talks at Bürgenstock, Switzerland, agreeing on a roadmap toward a final deal within 60 days, building on a 14-point memorandum of understanding signed on 17 June to de-escalate the war that began in February. The parties created working groups on the nuclear programme, sanctions, and monitoring and dispute resolution; Iran agreed to invite IAEA inspectors back into the country; and they established a contact channel to avoid incidents in the Strait of Hormuz and a "deconfliction cell" for Lebanon. Pakistan and Qatar mediated. The process remains fragile: in the same week, Israel killed scores of people in Lebanon despite the ceasefire there.
Sources: Al Jazeera (60-day roadmap, 14-point MoU of 17 June, working groups, IAEA inspectors, Hormuz channel), CNBC (roadmap, Lebanon deconfliction cell), Swiss Federal Department of Foreign Affairs (MoU between the USA and Iran).
Transparency notes
- Every claim above was checked against a primary or named source before drafting. The lead point's key facts (the 26 June partial reversal, the vetted-company list, and OpenAI agreeing to government vetting) come from an NPR interview with reporting based on a letter NPR has seen. The "roughly 100 companies and federal agencies" figure appears in CNBC's report; the script uses NPR's more conservative wording ("a small, named list of American companies").
- Editorial reads, not sourced facts: the ordering by long-term impact is an editorial judgement, not a sourced ranking. Framing points 1 and 2 as the US and EU "moving in opposite directions" is interpretation; the underlying dates and actions are sourced. Calling the Iran roadmap "fragile" is supported by the same-week Lebanon strikes but is a characterisation.
- Cut from the draft: several widely-covered items were dropped as market or personnel churn rather than durable structural shifts: ChatGPT falling below 50% of the assistant market (Gemini ~27.7%, Claude ~10.3%); Noam Shazeer moving to OpenAI and Nobel laureate John Jumper moving to Anthropic; SpaceX's S-1 disclosing a $1.25bn/month, 300MW compute deal with Anthropic; and Microsoft's 2026 capex guidance of ~$190bn. The Venezuela earthquakes (160+ dead) and the Ras Laffan, Qatar explosion (13 dead) were treated as tragedies without a clear six-month structural consequence and left out.