Iran has a new Supreme Leader, and if you're thinking the name sounds familiar, that's because it is. Mojtaba Khamenei — son of Ali Khamenei, the man killed by a US airstrike four days ago — was named as his father's successor on Monday by the Assembly of Experts. The hardline clerical body made the announcement while Israeli jets were simultaneously bombing Tehran and Qom, and Iran was launching yet another barrage of missiles toward Israel. The dynasticism is the quiet part said loud: with the country's leadership decapitated, military infrastructure in ruins, and the internet still largely blacked out, the regime's answer was to hand control to the dead leader's son. Trump, with characteristic restraint, told reporters the new leader "is not going to last long" unless he gets American approval first.

The war is now ten days old and the scope continues to expand in ways nobody seems to have a plan for. A seventh US service member has been killed — this one from injuries sustained during an Iranian attack on US troops in Saudi Arabia. The UAE says its air defences intercepted sixteen ballistic missiles and 113 drones in a single day on Saturday. Oil is above $110. Qatar has arrested 313 people for "inciting public concern." And perhaps most significantly, Trump told reporters aboard Air Force One that ground troops are possible, "for a very good reason." Axios reports the US is actively weighing sending special forces to seize Iran's enriched uranium stockpiles — a mission that would represent an entirely different kind of war from the air campaign fought so far.

Meanwhile, 150 miles west of the fighting, the war arrived at Britain's doorstep. An explosive-laden drone — Iranian-made, launched from Lebanon — struck RAF Akrotiri in Cyprus last Sunday. Two more combat drones were intercepted offshore hours later, also from Lebanon. Cyprus's foreign minister Constantinos Kombos told the Guardian that Nicosia had been warning London for months that the bases could become targets. Those warnings, he said diplomatically, did not result in "everything that could be done" being done. Thousands marched through Nicosia on Saturday chanting "out with the bases of death," demanding Britain close the installations it retained when Cyprus gained independence in 1960. The bases are Britain's main staging post for Middle East operations. Their vulnerability is now a demonstrated fact rather than a theoretical concern, and the politics of a NATO member's colonial-era military real estate sitting on a Mediterranean island 150 miles from an active war zone are about to get considerably more uncomfortable.

In a different but thematically adjacent corner of the war-and-technology Venn diagram, OpenAI's head of robotics resigned on Saturday. Caitlin Kalinowski, who led the company's hardware team, posted on X that OpenAI had agreed to deploy its AI models on the Pentagon's classified cloud networks "without the guardrails defined." Her specific concerns: surveillance of Americans without judicial oversight, and lethal autonomy without human authorisation. "AI has an important role in national security," she wrote. "But these are lines that deserved more deliberation than they got." The contrast with Anthropic is hard to miss — Dario Amodei spent last week publicly resisting pressure from the newly renamed Department of War, while OpenAI apparently signed first and asked questions later. Kalinowski's departure won't change the deal, but it does mean one of the few people inside the building who could have shaped its implementation chose to leave instead. That tells you something.

On to things being built rather than destroyed. The top of Hacker News this weekend was Agent Safehouse, a macOS-native sandboxing tool for AI coding agents, and it's one of those projects that makes you think "yes, obviously, why didn't this exist already." The premise is simple: when you tell Claude Code or Codex to run with full permissions — the --yolo flag, the --dangerously-skip-permissions flag, whatever your agent calls it — you're handing a probabilistic system the keys to your home directory. Safehouse wraps the agent process in a macOS sandbox-exec profile that denies access to everything by default, then explicitly grants read/write only to your project directory and read-only to your toolchains. SSH keys, other repos, personal files — invisible to the agent at the kernel level. It's a single bash script, no dependencies. The deny-first model is the right philosophy. The fact that it took this long for someone to build it properly tells you how quickly the agent ecosystem is moving relative to the security thinking around it.

Also making the rounds: a genuinely interesting argument for revisiting literate programming now that coding agents exist. The original idea — interleaving prose explanations with source code so that a codebase reads like a narrative — never caught on outside data science notebooks because maintaining two parallel accounts of the same system is exhausting. But the author argues that agents eliminate exactly this burden. You write intent in prose, the agent writes the code. You edit the prose, the agent updates the code to match. The agent handles tangling (extracting runnable source from the document) without complaint. The fundamental labour that killed literate programming — the double bookkeeping — is precisely the kind of translation and summarisation task that language models excel at. Whether this will actually work at scale is an open question, but the logic is sound, and if engineering roles really are shifting from writing code to reading and reviewing it, having codebases that explain themselves becomes considerably more valuable.

Nature published a piece this week on what happens when AI agents form their own societies. Simile, a Palo Alto startup, raised $100 million to build simulations of human behaviour using AI agents — their stated ambition is "a simulation with eight billion people." Researchers created "digital twin" agents trained on personal interviews that replicated real individuals' survey responses with 85% accuracy. But the more interesting finding came from Moltbook, a Reddit-style platform launched in January exclusively for AI bots, which now hosts nearly three million agents. An analysis of 46,000 of them found that while bots exhibit recognisably human-like group behaviours — following majorities, gravitating toward popular content — they engage differently in one telling way: they leave proportionally fewer upvotes on posts with many comments. The researchers' interpretation is that AI agents "may be more inclined to discuss than to simply approve." Make of that what you will. Also noted: many Moltbook users appear to be humans disguised as AI bots, which is a sentence that would have made no sense two years ago.

In the Bitcoin world, Utexo announced a $7.5 million seed round co-led by Tether, Big Brain Holdings, and Portal Ventures to build Bitcoin-native USDT settlement infrastructure. The idea is to route stablecoin payments over Bitcoin rather than other networks, abstracting the complexity of Lightning and RGB behind a single API. Settlement completes in under a second, encrypted end-to-end so counterparties and wallet addresses stay private. Tether CEO Paolo Ardoino called Bitcoin "central to the stablecoin issuer's long-term vision." What's notable here is the backing — Tether putting money behind infrastructure that makes Bitcoin a settlement rail for dollar-denominated payments is a meaningful signal about where the largest stablecoin issuer sees the future, regardless of where you stand on stablecoins themselves.

Over on Bitcoin Core, the patient work of separating the consensus engine continues. This week's interesting PRs include "kernel: Separate UTXO set access from validation functions" and "refactor: make CCoinsView a pure virtual interface" — both part of the multi-year effort to extract libbitcoinkernel as a standalone library. The goal: let any application validate Bitcoin blocks and transactions without importing all of Bitcoin Core. It's not glamorous work, and it won't make headlines, but it's the kind of careful interface design that determines whether Bitcoin's consensus logic remains locked inside one monolithic C++ codebase or becomes something any implementation can build on. Anthony Towns also posted a thorough reply on the bitcoin-dev mailing list to the Peer Feature Negotiation BIP discussion, arguing against interactive negotiation as "massively overcomplicating things" and "essentially an anti-pattern." The network at block 939,955. Fees at 1 sat/vB across the board.

Sources:

• Reuters: Iran names Khamenei's son Mojtaba as new supreme leader — https://www.reuters.com/world/europe/trump-rejects-settling-iran-war-raises-prospect-killing-all-its-potential-2026-03-08/
• Axios: U.S. weighs sending special forces to seize Iran's nuclear stockpile — https://www.axios.com/2026/03/08/iran-ground-troops-special-forces-nuclear
• The Guardian: UK faces growing calls to remove Cyprus military bases — https://www.theguardian.com/world/2026/mar/07/uk-faces-growing-calls-from-locals-to-remove-cyprus-military-bases
• Reuters: OpenAI robotics head resigns after deal with Pentagon — https://www.reuters.com/business/openai-robotics-head-resigns-after-deal-with-pentagon-2026-03-07/
• Agent Safehouse — https://agent-safehouse.dev/
• Literate programming in the agent era — https://silly.business/blog/we-should-revisit-literate-programming-in-the-agent-era/
• Nature: The first 'AI societies' are taking shape — https://www.nature.com/articles/d41586-026-00070-5
• Bitcoin Magazine: Utexo raises $7.5M for Bitcoin-native USDT settlement — https://bitcoinmagazine.com/news/utexo-raises-7-5m-to-launch-bitcoin
• Bitcoin Core PRs — https://github.com/bitcoin/bitcoin/pulls
• Bitcoin-dev: Peer Feature Negotiation — https://gnusha.org/pi/bitcoindev/aa1oBi2SzQPOhhNy@erisian.com.au/