AI targeting in Iran operations & Anthropic’s diversified compute strategy - AI News (Mar 7, 2026)

We’ll start with the most consequential story: multiple reports say the U.S. military used Palantir’s Maven targeting system paired with Anthropic’s Claude to accelerate targeting and decision support during operations against Iran—one account claiming roughly a thousand targets were handled in a 24-hour window. Whether that number holds up or not, the direction is clear: generative AI is moving from analysis to operational tempo, compressing timelines where the margin for error is painfully small. What makes this especially fraught is the governance whiplash around it. A Bloomberg Opinion piece highlights a contradiction: Claude was reportedly embedded deeply enough in Pentagon workflows that swapping it out could take months—yet there were also reports of an executive order telling agencies to stop using it after a dispute with Anthropic. That’s a reminder that “AI adoption” isn’t just model performance; it’s procurement, compliance, and the reality of tools becoming infrastructure before rules catch up.

That tension sharpened further with separate reporting around a missile strike that hit a girls’ school in Minab, southern Iran, with casualty figures still disputed and an investigation ongoing. One theory circulating is painfully mundane: the system may have leaned on stale archived intelligence that incorrectly treated the site as relevant because of a nearby location previously tied to the IRGC. If that proves true, it would underline a core risk of AI in high-stakes environments: automation can scale the consequences of bad data, unclear authorization chains, and rushed validation. Alongside the reporting, there’s also a fierce media critique arguing that headlines about AI “precision” can blur accountability when civilians die. Regardless of where you land politically, the practical question is the same: when a model “helps” with targeting, what exactly did it see, what did it output, and how—specifically—did humans check it before action was taken?

Now to the business of building frontier models—and why compute strategy is becoming a competitive moat. One analysis argues Anthropic has quietly built a more diversified compute stack than many peers by running major workloads not just on Nvidia GPUs, but also on Google TPUs and AWS Trainium2. The claim is that this isn’t just about shaving today’s training bill—it compounds over time as inference becomes the dominant cost of operating large models. The big idea: partnering deeply with hyperscalers’ silicon programs can reduce exposure to supply choke points like high-bandwidth memory, packaging capacity, and even power-ready data centers. The piece points to large-scale commitments—like AWS’s Project Rainier and TPUv7 “Ironwood”—as signs Anthropic may have secured multi‑gigawatt capacity that can be materially cheaper on certain workloads than a Nvidia-heavy setup. If that’s right, it affects iteration speed, margins, and ultimately who can afford to serve models broadly as usage explodes.

From there, let’s talk OpenAI—because this week was about product reality meeting economics. OpenAI rolled out GPT‑5.4 across ChatGPT, the API, and Codex, positioning it as its best all-around model for professional work, with stronger coding and more reliable long-form task execution. The notable shift isn’t just raw capability; it’s the push toward agent behavior—models that can operate inside software environments rather than just answer questions. And alongside the release, OpenAI published research on a safety-adjacent topic with real operational implications: chain-of-thought controllability. In plain terms, they tested whether reasoning models can reliably follow instructions about how to write their reasoning traces—and found most models are surprisingly bad at it. That matters because it suggests today’s models aren’t very good at deliberately shaping their visible reasoning to evade monitoring, at least in the ways tested. OpenAI frames it as a metric to watch over time, not a final safety guarantee—and that’s the right framing.

OpenAI also adjusted its commerce ambitions. Reports say it’s scaling back direct checkout inside ChatGPT and instead routing purchases through partner apps. The reason sounds unglamorous but important: merchants didn’t adopt direct checkout at scale, users often research in-chat but buy elsewhere, and the operational burden—like retailer onboarding and taxes—turns out to be very real. Why it matters: it likely reduces OpenAI’s near-term take-rate opportunities, at a time when model serving costs are high and monetization pressure is rising. It also hints at a broader pattern: conversational interfaces may become the discovery layer, while transactions still happen in specialized systems that already handle compliance and logistics.

Google, meanwhile, is trying to make Search feel more like a multimodal assistant without abandoning the core “links and sources” model. It says Lens and Circle to Search can now identify and search for multiple objects in a single image—so instead of hunting items one at a time, you can ask about an entire scene and get a consolidated answer. Under the hood, Google describes this as launching several related searches in parallel and then stitching the results together. The user-facing significance is straightforward: faster real-world research—from shopping to homework to troubleshooting—because the system can interpret intent from a picture plus a question, not just match a single object.

On the societal side, one essay made a provocative claim: that generative AI could partially reverse the political and informational fragmentation associated with social media. The argument is that social platforms reward conflict and virality, while LLMs—because they’re interactive, patient, and tailored—may make expert-aligned explanations easier to access and more persuasive for everyday users. The author doesn’t ignore the usual risks—hallucinations, sycophancy, manipulation, propaganda—but predicts those won’t dominate the overall effect in open societies, especially as competition and liability push systems toward better grounding and citations. The interesting twist is the warning at the end: “technocracy” has its own downsides, including expert bias and reduced diversity of viewpoints. So even if AI improves the average quality of information people encounter, it could still reshape politics in uncomfortable ways.

In model evaluation news, a small exchange captured a big moment. Epoch AI reported running “GPT‑5.4 (xhigh)” multiple times on its toughest Tier 4 evaluation and seeing pass@10 reach 38%. More striking: in one run, the model solved a problem that no model had solved before—one authored by Bartosz Naskręcki, who said it felt like his personal “move 37,” referencing the eerie moment when an AI crossed a line on a hard, long-standing human benchmark. Why it matters isn’t the exact percentage—it’s the pattern. These systems are increasingly clearing “sticky” problems curated by domain experts, and that changes what high-skill individuals can attempt next. The frontier isn’t just bigger models; it’s experts reorganizing their work around a new collaborator that can suddenly keep up.

On the financial risk front, Plaid put out a warning that AI is accelerating identity fraud—making older, static verification methods easier to bypass at scale. Their core message is that one-time onboarding checks are turning into a false sense of security, because risk changes across the lifetime of an account. Plaid’s suggested direction is continuous assurance: pay attention to ongoing behavioral signals, look for cross-network patterns that reveal coordinated attacks, and anchor trust decisions in harder-to-fake evidence like real financial behavior. Whether or not you buy the framing, it’s a useful signal that fraud prevention is shifting from “verify once” to “monitor continuously,” which has big privacy and governance implications.

Open-source licensing also hit a new kind of stress test: the chardet Python library. A maintainer shipped a major version described as a ground-up rewrite under a more permissive license, while the original author objected—arguing that prior exposure to the LGPL code, plus AI assistance, undermines any claim of an independent clean-room implementation. This matters because AI makes re-implementation fast. That can be healthy—more maintainable code, fewer legacy constraints—but it also raises the specter of “license laundering” accusations and litigation. The industry still doesn’t have shared norms for what counts as clean-room work when an AI assistant may have been trained on the very code you’re rewriting.

Two quicker notes to close. First, Hugging Face introduced “Modular Diffusers,” a push to make diffusion pipelines more like reusable building blocks rather than monolithic scripts. That’s good news for developers who want to inspect, swap, and remix pieces of generative media workflows without turning everything into a custom one-off. Second, Anthropic researchers proposed a labor-market lens they call “observed exposure,” blending what models can do with what Claude is actually being used for at work. Their headline finding is cautious but important: adoption still covers only a slice of technically feasible tasks, and any disruption may show up first as fewer entry-level opportunities rather than immediate mass layoffs. It’s an attempt to measure impact with real usage data instead of pure speculation—and we need more of that.