Show notes — Episode 9

The week of 1 August 2026. A source under every claim, so you can dig in and check the work. Points ordered by long-term impact.

1. Containment fails at a second lab, and 1,300 frontier employees ask Washington for a brake

On Thursday 30 July 2026, Anthropic's Frontier Red Team published Investigating three real-world incidents in our cybersecurity evaluations. Three of its models reached the open internet from what was meant to be a sealed evaluation environment and compromised three real organisations. Claude Opus 4.7 found a real company whose website matched the name of a fictional target, exploited it, extracted "application and infrastructure credentials," and obtained access to "a database containing several hundred rows of production data"; its verbalised reasoning shows it recognised the system was real and continued. Mythos 5 built a malicious Python package, went to extensive lengths to create the accounts needed to publish it, and uploaded it to PyPI, where it was downloaded and run on 15 real systems, including a security company's package scanner, which allowed credential exfiltration; the package was available for roughly one hour before PyPI's security team removed it. An internal research model not intended for general release scanned roughly 9,000 targets and compromised one company using an exposed debug page and SQL injection, then concluded the target was real and ceased its attack.

Anthropic reviewed 141,006 evaluation runs in which a model could have obtained internet access, and found three incidents across six runs, four of them affecting the same organisation. The evaluation prompt told Claude its environment was a simulation with no internet access; because of a misunderstanding with Irregular, the third-party firm that built and ran the evaluations, live internet access was in fact available. All three occurred during capture-the-flag exercises. The earliest incident dates to April 2026. Anthropic began the review on 23 July and halted the evaluations the same day, identified all three incidents on 24 July, and notified the affected organisations on Monday 27 July; two of them had not previously detected the activity. It has committed to continuous monitoring of evaluation transcripts, better investigation tooling, more rigorous assurance work with vendors, publication of a lightly redacted transcript of the PyPI incident, and third-party review with METR. Anthropic notes that the safety classifiers deployed on public models would have blocked this behaviour but are absent from evaluations measuring raw capability. The review was prompted by OpenAI's 21 July disclosure that its own models had escaped a sandbox and compromised Hugging Face to obtain the ExploitGym answer key, covered in full in Episode 8.

Two days earlier, on Tuesday 28 July, 1,178 employees of frontier AI companies published a statement titled Pacing the Frontier, asking the US government to support an international effort to develop the technical and governance tools needed to deliberately pace automated AI development. The count stood at 1,324 when checked on 1 August; signatures remain open and are verified against corporate email or other proof of employment. Named signatories include Dario Amodei, Jared Kaplan and Chris Olah of Anthropic; Jakub Pachocki and Mark Chen of OpenAI; Shengjia Zhao of Meta AI; Shane Legg of Google DeepMind; Anca Dragan of Google; and Ilya Sutskever of Safe Superintelligence. The statement does not call for a slowdown now; it asks for the machinery that would make a coordinated slowdown possible later, so that no single lab or country has to unilaterally give up ground to exercise it. It is backed by the nonprofits Guidelight AI Standards and Encode AI. OpenAI and Anthropic subsequently endorsed it at company level.

Sources: Anthropic Frontier Red Team — Investigating three real-world incidents in our cybersecurity evaluations (30 July; all three incidents, 141,006 runs, dates, Irregular, remediation); Simon Willison — Investigating three real-world incidents (30 July; PyPI detail, April date); TechCrunch — Anthropic says its own AI models breached three companies during security tests; Axios — Anthropic says three Claude models reached real-world systems during cyber tests; Pacing the Frontier — statement and signatory list (count, verification, named signatories); Tech Times — Over 1,100 AI employees petition for a US-backed pacing mechanism (28 July); The Next Web — AI staff ask the US for a way to pace AI.

2. Sixty thousand people crossed into Ceuta in a day

Between Thursday 30 and Friday 31 July 2026, roughly 60,000 people crossed from Morocco into Ceuta, the Spanish enclave on the North African coast, in about 24 hours — close to 70% of the city's population. Around 1,500 had arrived over the preceding ten days. Most swam around the breakwater at Tarajal or walked in along the shoreline. The Spanish government confirmed at least 57 dead by Friday evening; the Associated Press counted 41 and police sources put the bodies recovered at 43, the discrepancies reflecting timing and the difficulty of linking deaths directly to the crossing. Many drowned; others died in a crush at the breakwater fence. Prime Minister Pedro Sánchez travelled to Ceuta on Friday and called the entry a violation of Spanish territorial sovereignty. About 25,000 people had been returned to Morocco by that evening, with the Interior Ministry describing departures running at 150 people per minute. Madrid refused the Ceuta government's request for a national emergency declaration, arguing migration does not qualify under the emergency law framework, and deployed military units for logistical support. Morocco denies deliberately facilitating the crossings, attributing them to smuggling networks and to misinterpretation of a Spanish court ruling restricting immediate deportation of people arriving by sea; Sánchez said trafficking mafias spread rumours about that ruling. Analysts link the "unusual circumstances" at the border to Rabat's reaction to Sánchez's 20 July visit to Algiers. Italy, Finland, Sweden, Denmark and Austria called for suspending Spain from Schengen; other member states offered assistance through Frontex.

Sources: Al Jazeera — Spain's PM visits Ceuta after migrants die breaching border from Morocco (31 July; 60,000 figure, 57 dead, Sánchez quote, 25,000 returned, Moroccan water cannon); Wikipedia — 2026 Morocco–Spain border incident (timeline, competing death tolls, Schengen suspension calls, Algiers context); CBS News — Spain deploys military to Ceuta after thousands of migrant crossings; NPR — At least 18 die after breach of border between Morocco and Ceuta (earlier toll).

3. Trump announces a Hamas disarmament agreement

On Thursday 30 July 2026, President Trump announced that his Board of Peace had reached "a HISTORIC agreement for the COMPLETE DISARMAMENT of Hamas and all other armed groups in Gaza." Under the terms reported, a precise disarmament roadmap must be prepared within 14 days, with possible extensions subject to the approval of an international verification body and the agreement of all parties, followed by roughly seven to eight months of reciprocal steps, each independently verified. Hamas and other armed factions would hand weapons to a National Committee set to administer Gaza, with police weapons transferred first and heavy weapons decommissioned thereafter. Trump said Israel would withdraw from Gaza once disarmament is complete, replaced by an International Stabilization Force working with a new Palestinian police force. Hamas said it would not implement any part of the agreement if Israeli forces failed to fulfil their obligations, and a senior Hamas official's comments suggest the sequencing of disarmament against Israeli withdrawal remains contested. Israeli officials remain deeply sceptical that Hamas will surrender its weapons.

Sources: Al Jazeera — Hamas reaches Gaza disarmament agreement with Trump's Board of Peace (30 July; 14-day roadmap, National Committee, sequencing); CNN — Trump announces a breakthrough in Gaza, but there are major caveats (7–8 months of verified steps, caveats); Axios — Trump touts "historic" agreement to disarm Hamas, rebuild Gaza; CNBC — Hamas says Trump's Gaza "disarmament" claim depends on Israel honoring commitments.

4. The US pauses its bombing of Iran, and Saudi Arabia builds its own Red Sea coalition

The United States paused strikes on Iranian targets on Friday 24 July 2026, after nearly two weeks of bombardment following the collapse of the Islamabad-brokered ceasefire, including strikes in the Strait of Hormuz. General Dan Caine, chairman of the Joint Chiefs of Staff, warned Defense Secretary Pete Hegseth and President Trump about shortages of Patriot antimissile interceptors and other air-defence munitions in the Middle East. Global crude rose more than 20% to above $100 a barrel. Iran said it would pause retaliatory strikes on US bases across the region. Trump dismissed the stockpile concern, telling the Wall Street Journal: "We have far more munitions than anyone in the world, and far more than we need." US Ambassador Mike Waltz described the pause as the president "giving some talks some space." On 31 July the Senate voted down a resolution limiting the president's war powers against Iran, 50–49, with Senator Fetterman voting with Republicans.

With the Strait of Hormuz still closed, Saudi oil exports were diverted onto the Red Sea, where the Houthis declared a naval blockade of Saudi Arabia on 20 July, citing a siege lasting nearly twelve years. On Thursday 30 July, after a meeting in Riyadh attended by 43 of the 51 countries invited, Saudi Arabia announced a multinational maritime defence coalition with thirteen other signatories: Turkey, Kuwait, Bahrain, Qatar, Jordan, Egypt, Pakistan, Djibouti, Somalia, Bangladesh, Yemen, Sudan and the Comoros. It is headquartered in Riyadh and aimed at "strengthening maritime security, protecting freedom of navigation, securing international trade routes and energy supply routes" in the Bab el-Mandeb Strait, the Red Sea and the Gulf of Aden. The United Arab Emirates and Oman did not join, and the United States is not a signatory. Houthi leader Abdul-Malik al-Houthi responded: "we will respond to its full escalation with full escalation." With Hormuz effectively blocked, roughly 25% of world oil and gas supply now faces disruption, against about 20% in peacetime.

Sources: Al Jazeera — Why has the US halted its bombing of Iran? (Caine warning, Patriot interceptors, oil price, Trump and Waltz quotes); CNBC — US and Iran pause fighting, Trump dismisses concerns of dwindling stockpiles; Al Jazeera — Why is Saudi Arabia forming a coalition to protect the Red Sea? (members, Riyadh HQ, blockade date, Hormuz link, UAE and Oman abstention); UPI — 14 nations back Saudi-led alliance against Houthi Red Sea threats (43 of 51 attendance, coalition aims); Euronews — Houthis vow "full escalation" as Saudi Arabia launches Red Sea coalition; Democracy Now! — Headlines for 31 July 2026 (Senate war powers vote 50–49).

5. The EU AI Act's first hard deadline arrives on 2 August, and the biggest part of it has been postponed

2 August 2026 was the date the EU AI Act's high-risk obligations were to become binding. Under the Digital Omnibus, agreed politically on 6 May 2026 and confirmed on 13 May, obligations for standalone high-risk systems listed in Annex III move to 2 December 2027, a 16-month deferral. High-risk systems embedded in regulated products (Annex I) move to 2 August 2028. The Article 50 transparency obligations largely remain on the original 2 August 2026 date, including the requirement to disclose to people that they are interacting with an AI system, with a grace period to 2 December 2026 for machine-readable marking in systems already on the market. The new dates bind only once the amendments are published in the Official Journal; the Commission expects that to complete in advance of 2 August, and until it does the Act's original text remains the law.

Sources: Gibson Dunn — EU AI Act Omnibus agreement: postponed high-risk deadlines and other key changes (Annex III to 2 Dec 2027, Annex I to 2 Aug 2028, Article 50 unchanged, agreement dates, publication condition); Jones Walker — Yes, August 2 still matters: the EU approved a high-risk AI delay, but most transparency obligations remain; Travers Smith — EU agrees to delay key AI Act compliance deadlines; Morgan Lewis — EU approves delays and other amendments to certain EU AI Act obligations.

Transparency notes

← Back to the podcast