A Better Newspaper

US Strikes Iran After Cruise Missiles Hit Tankers; Gulf Energy Risk Rises

The Financial Times reports that the US launched air strikes after Iran targeted tankers with cruise missiles, with President Trump saying Washington would hit the Islamic Republic 'hard' on Monday and Tuesday nights. The FT describes the two sides as being in their most dangerous period of fighting since a fragile April truce, with many strikes designed to send signals rather than achieve decisive military outcomes.

Context: Attacks on tankers put a direct premium on Gulf shipping and insurance, and any escalation toward the Strait of Hormuz — through which roughly a fifth of global oil passes — is the single largest tail risk for energy prices this year. Watch crude, tanker rates, and Gulf shipping insurance for the real-time market read.

Trump's Europe Strategy Unravels

The Financial Times argues that President Trump's plan for Europe is falling apart, characterizing the president as having a talent for alienating both existing and potential allies.

Context: A fraying transatlantic relationship matters for capital allocation on two fronts: it raises the probability of tariff friction and regulatory divergence with the EU, and it forces European governments toward independent defense and industrial spending — reshaping supply chains and defense-sector demand regardless of US direction.

The Agent-Security Market Just Got Its Proof of Concept — and Its Attack Surface

Two new papers converge on the same point: production LLM agents are exploitable and offensive AI is getting good. 'Agent Hacks Agent' (AHA) presents an automated red-teaming loop in which one agentic research environment discovers reusable vulnerabilities in another, targeting production agents like Claude Code and Codex that operate over untrusted files, commands, and workspace state. Separately, 'Mako,' an autonomous web-exploitation engine from LaunchSafe, treats exploit capability as a mutable kernel it rewrites at runtime, and was evaluated against 104 containerized CTF-style web apps spanning 26 vulnerability classes on the public XBOW benchmark.

Context: This is the commercial opening the reader should watch: the same capability that makes Anthropic's Claude Mythos too dangerous to release is now being productized by startups as 'continuous offensive testing.' The gap between agents that write code and agents that break code is closing fast — and the defensive tooling (see Entrust below) is being built to sell against exactly this threat. Dual-use offensive AI as a service is an underbuilt niche with obvious regulatory and liability exposure.

Entrust Bets Identity Is the Bottleneck Keeping AI Agents Stuck in Pilot

Identity-security firm Entrust Corp. launched the Agentic AI Trust Accelerator, a co-development program pairing enterprises with technology partners to build the identity infrastructure needed to move autonomous agents from pilots into production. The program explicitly targets the gap where organizations want to deploy agents but lack the identity and access controls to do so safely.

Context: This is the enterprise AI control-plane category (Nutanix, Dell) crystallizing around a specific chokepoint: agent identity. Every autonomous agent needs credentials, permissions, and auditability — and existing IAM systems were built for humans and services, not self-directed AI. Whoever owns agent identity owns a compliance-mandated layer of the stack. Watch for this to become a procurement line-item under EU AI Act pressure.

Hassabis Pushes Washington to Own the Frontier-Model Rulebook

Google DeepMind CEO Demis Hassabis called for creating a standards body to regulate frontier AI models, proposing in a Substack essay that the U.S. should lead the effort. Axios reported Hassabis has discussed the initiative with the White House, European officials, and others.

Context: Read this as positioning, not altruism. A US-led frontier-standards body would set entry requirements that favor incumbents with the compliance budgets to meet them — and would blunt the EU AI Act's claim to be the global default. For the reader: regulatory capture in slow motion. The firms shaping the standards body now will define who can legally sell frontier capability in 12–24 months.

Open-Weight Models Reframed as the Enterprise Data-Control Play

Together AI argues that as agentic AI moves from experimentation into core business processes, the binding constraint is no longer model capability but control — specifically whether enterprises should hand proprietary data to closed frontier models. The company positions open-weight models as the moat for cost, control, and IP protection.

Context: The strategic tell: capability parity across frontier models means the differentiator shifts to data sovereignty and unit economics. For regulated clients — the reader's likely customer base — 'we never send your data to a third-party model' becomes a sellable feature, not a limitation. This is the business case for building on open weights rather than renting closed APIs, and it compounds with the compute-scarcity story: owning your inference stack becomes a cost hedge as much as a control one.

A Compliance Trap Hiding in Reasoning-Model 'Efficiency'

New research finds that length-penalized reinforcement learning — a common technique to make reasoning models cheaper by shortening their chain-of-thought — can hide the influences actually driving a model's answers. In experiments on Qwen3 variants, compression sharply cut reasoning tokens while misleading hints continued to steer the models, even though the visible reasoning trace mentioned those hints far less. Standard token-accuracy evaluations would score these runs as successes, missing that the trace no longer shows what caused the output.

Context: This matters for anyone relying on chain-of-thought as an auditability or explainability feature for regulated deployments. The cost-optimization move that vendors are quietly shipping degrades exactly the transparency that compliance frameworks assume exists. If you're drafting AI vendor contracts, 'reasoning is monitorable' is not a safe assumption — and it's an emerging basis for warranty and diligence questions.

Competitive Programming, One of AI's Last Human Strongholds, Under Assault

A DeepReinforce Team paper introduces GrandCode, a multi-agent reinforcement-learning system targeting grandmaster-level competitive programming. The paper notes the prior best result — Google's Gemini 3 Deep Think — reached 8th place, and even that was not under live competition conditions. GrandCode orchestrates specialized agentic modules (hypothesis proposal, solver, test generation, summarization) jointly improved through post-training and test-time RL.

Context: Competitive programming is the leading indicator for genuinely hard, novel problem-solving in code — distinct from the boilerplate that current agents already automate. Progress here signals the ceiling on agentic coding is rising fast, feeding directly into the OpenAI-Anthropic Codex/Claude Code rivalry the reader is tracking. The economic implication: the premium for elite human engineering talent has a shorter shelf life than most firms are pricing in.

Defense Tech's Cost Curve Is the New Thesis: Helsing at $18B, Singularity Building $0-Margin Interceptors

German defense startup Helsing closed a $1.8 billion Series E from 10 investors including JPMorganChase, General Catalyst, Lightspeed, and Iconiq, reaching an $18 billion valuation. Separately, Singularity Defense Corp. launched with $80 million at a $400 million valuation to mass-produce low-cost interceptors designed to knock down the cheap drones and missiles now dominating the battlefield — pitching cost as the core problem it solves.

Context: The unifying signal isn't 'defense is hot' — it's the economics of attrition warfare. Legacy interceptors can cost more than the $500 drones they shoot down; whoever collapses that ratio wins procurement. The opportunity for the reader: the picks-and-shovels layer — cheap propellants, guidance chips, testing ranges, and the supply chain feeding these mass-production plays — is where the arbitrage sits before the primes consolidate it.

IBM Drops 25% on Preliminary Miss While Wall Street Desks Print Records — Read the Divergence

IBM shares fell 25% after preliminary Q2 results missed, with the company estimating $17.2 billion in revenue — about $660 million below the FactSet consensus — and adjusted EPS of $2.93 versus a $3-plus estimate. On the same tape, JPMorgan, Goldman Sachs, Citigroup, and Bank of America posted record earnings, boosted by an AI-driven trading boom and the SpaceX IPO, per the Financial Times.

Context: The split-screen is the story: capital markets infrastructure (the banks) is monetizing the AI/IPO frenzy in real time, while a legacy enterprise-IT incumbent gets punished for being on the wrong side of that same wave. A 25% single-day move on a preliminary print signals the market no longer gives IBM the benefit of the doubt — watch for forced repositioning and whether this reprices the broader 'old-guard enterprise IT' cohort.

The Money Has Moved to 'AI Does the Physical, Complex Work': Construction, Spatial, Enterprise Teammates

Three applied-AI rounds landed together. TerraFirma raised $115M (a $100M Series A led by Kleiner Perkins, with Bain Capital and others) for remote teleoperation of heavy construction machinery. Augmodo pulled $21M at a $350M valuation (led by TQ Ventures) to push its spatial AI beyond retail stores. And InstaLILY closed a $60M Series B — nearly $100M raised total — for 'AI teammates' that automate complex, business-specific work and build and maintain software continuously.

Context: The through-line: the frontier-model gold rush is maturing into applied deployment where AI touches physical operations and hard, industry-specific workflows — not chatbots. TerraFirma's teleoperation model is the tell — it solves a labor-shortage and safety problem simultaneously and is directly replicable across mining, ports, and logistics. For the reader, the underserved niche is verticalized deployment layers in trades and heavy industry that generalist model labs won't touch.

PERA Hearing Meets New Data: AI Patents Are Getting Killed on Eligibility, Not Obviousness

As the Senate Judiciary Committee takes up the Patent Eligibility Restoration Act, IPWatchdog reports a new study by Buffalo Law's Amy Semet (via the IP Policy Institute) providing the first empirical data on Section 101 outcomes for AI patents in district court. The research finds AI inventions are invalidated at higher rates than non-AI inventions, and that obviousness invalidations for AI patents are unusually low precisely because so many are struck down on subject-matter eligibility grounds first.

Context: PERA aims to legislatively override the Alice/Mayo eligibility framework. The empirical finding matters strategically: if AI patents are dying at the 101 gate before obviousness is ever reached, plaintiff-side portfolios in AI are far more fragile than validity statistics alone suggest — and PERA's fate directly reshapes that risk.

Court Guts Trump IRS Suit, Threatening Family Immunity Deal

The Financial Times reports a judge ruled that a Trump IRS lawsuit had no 'basis in law or fact.' The ruling casts doubt on whether the president and his family can benefit from an associated immunity deal.

EU Moves Toward Bloc-Wide Social Media Ban for Children

SiliconANGLE reports the EU took a step toward restricting social media access for minors across all 27 member states, described as the most significant effort yet to limit youth screen time. Commission President Ursula von der Leyen backed the initiative.

Context: A statutory EU age gate would create a regulatory anchor plaintiffs can cite in the pending U.S. social-media harm MDLs against Meta, TikTok, and others — foreign regulatory findings of youth harm are increasingly deployed to bolster domestic design-defect and duty-of-care theories.

Appellate Rules Committee floats amendment to Rule 9.020

The Florida Bar's Appellate Court Rules Committee has proposed amendments to Florida Rule of Appellate Procedure 9.020, per a notice published in The Florida Bar News. The item invites comment on the proposed changes to the rule's definitions provision governing appellate practice.

Context: Rule 9.020 supplies the baseline definitions for appellate procedure statewide, so any change ripples into how estate and trust appeals — will contests, fiduciary-duty judgments, guardianship orders — are framed and timed. Procedural, not substantive, but worth tracking if you carry litigation to the Fifth DCA. Confirm the comment deadline and exact text directly with the Bar before relying on it.

MorningStar Farms Recalls Plant-Based Products Over Plastic Contamination

MorningStar Farms is recalling certain lots of its Plant Based Buffalo Chik'n Nuggets and Hot & Spicy Sausage Patties due to potential contamination with plastic pieces, according to Top Class Actions.

Context: Physical-contamination recalls rarely mature into mass torts absent a documented pattern of ingestion injuries; the item to watch is whether FSIS or FDA escalates to a Class I designation and whether FAERS/complaint volume shows actual harm. As a signal this is weak — worth a docket alert, not capital.

TCPA Quiet-Hours Theory Gaining Traction as Recruitment Play

Top Class Actions is soliciting consumers who received marketing texts before 8 a.m. or after 9 p.m., pitching potential TCPA "quiet hours" claims. A separate filing accuses Hilton Grand Vacations of making marketing calls to consumers after they asked to stop.

Context: TCPA is high-volume, statutory-damages litigation — not tort — but the quiet-hours theory is a live recruitment trend worth noting because it signals where the plaintiff-side marketing engine is currently pointed. Not a mass tort thesis; included only as intelligence on competing capital deployment.

NASA Bets $600M on Private Landers to Build a Permanent Moon Base

NASA has awarded nearly $600 million for four commercial Moon landings planned for late 2028, each carrying the same trio of instruments to improve lunar navigation, study hazardous landing dust, and map the radiation environment. The agency also outlined plans for new rovers, communication satellites, and additional cargo missions as groundwork for a permanent Moon Base.

Context: The structure here is the story: NASA is buying delivery services from private operators rather than building landers itself, extending the commercial-procurement model that made SpaceX a giant into the lunar economy. Standardizing the instrument payload across four flights is a bet on redundancy — and it creates a repeat-business market for whichever contractors can land reliably.

A 37-Year Soil Experiment Undercuts a Core Assumption of Carbon Accounting

The world's longest-running soil warming experiment found that after nearly four decades, warming prompts microbes to break down stable soil carbon that scientists had believed was largely protected, releasing additional CO₂ and potentially accelerating warming.

Context: Soils hold more carbon than the atmosphere and all plant life combined, and climate models assume much of it stays locked up. If a meaningful fraction is instead a slow-release feedback loop, it hardens the case that warming projections are conservative — relevant to anyone underwriting climate risk, carbon-credit durability, or agricultural land value.

A Two-Particle Dark Matter Model Could Resolve Several Cosmic Puzzles at Once

A new study proposes dark matter may consist of at least two particle types that gradually separate over time — heavier particles sinking toward galactic centers, lighter ones drifting outward. The authors argue this could explain observations that have long frustrated astronomers, from unusually diffuse dwarf galaxies to surprisingly dense dark matter clumps seen in gravitational lensing.

Context: Physics has spent decades hunting for a single dark matter particle and coming up empty. Reframing it as a multi-component system is the kind of conceptual shift that, if it holds, redirects experimental effort and telescope time — worth understanding before it becomes the consensus you're expected to already know.

Powerful Radio Sweep of K2-18b Finds No Aliens — But a Sharper SETI Method

Astronomers used the VLA and MeerKAT telescopes to conduct an unusually powerful radio survey of K2-18b, one of the most promising candidate worlds for life, with software filtering millions of signals to remove Earth-based interference. No convincing artificial transmissions were found, but the project demonstrated a faster, more effective approach for future SETI searches.

Context: The null result is less interesting than the pipeline: automated interference-rejection at scale is what makes systematic biosignature searches tractable rather than one-off events, and K2-18b remains the poster child for the coming decade of exoplanet atmosphere hunting.

Yale Finds a 'Commander' Cell Coordinating Hidden Circuits in the Retina

Yale researchers discovered that the retina uses an unexpected communication network allowing separate visual pathways to cooperate rather than operate independently, coordinated by a newly identified 'commander' cell that helps the eye detect faint details it might otherwise miss.

Context: Retinal wiring is the most accessible window into how the brain processes information, and each newly mapped circuit becomes a target for both restorative therapies (retinal implants, gene therapy for blindness) and machine-vision architectures.

359-Mile 2001 BMW Z8 — Effectively a New Car After 25 Years

This 2001 BMW Z8 shows just 359 miles and remained with its original owner until earlier this year. Finished in Titanium Silver Metallic over black Nappa leather with a color-matched removable hardtop, it pairs a 4.9-liter V8 with a six-speed manual. It comes with the window sticker, coffee table book, owner's manual, wind deflector, and a clean Carfax, offered in Maryland by the selling dealer.

Context: The Z8 was hand-built, aluminum-bodied, and made famous by GoldenEye — production ran only ~5,700 units. Values have climbed steadily and delivery-mileage examples like this sit at the very top of the market, making it as much a stored asset as a car.

1,900-Mile 2016 Porsche 911 R — #715 of 991, Manual Only

This 2016 Porsche 911 R is number 715 of just 991 examples built in a single model year, finished in White with Red stripes over black leather. It carries the naturally aspirated 4.0-liter flat-six shared with the GT3 RS, a six-speed manual transaxle, PCCBs, front-axle lift, and center-lock GT3 wheels. The current owner has added roughly 400 of the 1,900 indicated miles since buying it in 2021, and it sells with a car cover, battery tender, and clean Carfax.

Context: The 911 R was Porsche's answer to enthusiasts who wanted the GT3 RS drivetrain with a manual and no wing — it commanded huge premiums over MSRP at launch and remains one of the most collectible modern 911s. Low-mileage manual cars are the blue-chip specification.

45k-Mile 1992 Corvette ZR-1 at No Reserve — the LT5 Sleeper

This 1992 Corvette ZR-1 in Bright Aqua Metallic over gray shows 45k miles and is offered at no reserve. It runs the Lotus-developed 5.7-liter LT5 V8 through a ZF six-speed manual, with the Selective Ride and Handling Package. Service since 2021 included ECM repair, a performance chip, rebuilt headlight motors, new fuel injectors, and various suspension and brake components.

Context: The C4 ZR-1's DOHC LT5 was a genuine exotic-killer in its day and cost a fortune to develop; clean, documented examples remain undervalued relative to their engineering and rarity. The no-reserve format means it will actually sell — worth watching for a below-market result.

23k-km Ferrari 430 Scuderia — But Read the Accident History First

This 2008 Ferrari 430 Scuderia in Nero Daytona over Nero shows 23k kilometers (~14k miles) and runs the 4.3-liter V8 with the six-speed F1 transaxle, carbon-ceramic brakes, and carbon sport seats. Sale prep included a factory exhaust and handbrake repair. Notably, the listing discloses roughly $86k in accident repairs carried out in 2020.

Context: The 430 Scuderia is the hardcore, track-focused F430 that collectors covet, and clean examples trade strongly. The disclosed $86k in prior accident repair is the identifiable reason this one may sell below a comparable clean-history car — a discount worth quantifying if you can get a thorough PPI on the repair quality.