A federal deadline lands today that changes how the most capable AI models reach the market, behind a classification barrier. Elsewhere: DeepSeek pushed the price floor for agentic coding down again, Amazon raised its 2026 infrastructure budget to $220 billion while saying it still cannot meet demand, new research put a number on hidden prompt injections in job applications, a German court ruled against a generative music company, and a study of 175 brands found AI assistants recognize almost everyone and recommend almost no one.
The government's frontier model framework is due today, and the benchmark is classified
Today marks the 60-day deadline under the executive order President Trump signed on June 2, "Promoting Advanced Artificial Intelligence Innovation and Security," known as Executive Order 14409. By August 1 the Treasury Department, the NSA, and CISA were required to develop a classified benchmarking process for assessing the advanced cyber capabilities of AI models, establish the threshold defining a "covered frontier model," and design the voluntary framework governing what happens once a model crosses it. NIST, the National Cyber Director, and the President's science adviser consult; the NSA Director makes the designations.
The framework is voluntary in a narrow sense. A developer may work with the government to determine whether a model meets the threshold, may give agencies access for up to 30 days before public release subject to confidentiality and IP protections, and may help select which "trusted partners" get early access. The order expressly disclaims any mandatory licensing, preclearance, or permitting requirement. Nothing becomes binding on any company today — August 1 is a deadline for the government, not for industry.
The part worth watching is the classification. Because the NSA-led criteria are classified, developers will not see where the line sits: a lab can be told a model qualifies as covered, but it cannot engineer around a threshold it cannot read. The same order directs the Attorney General to prioritize prosecutions under the identity fraud, Computer Fraud and Abuse Act, and wire fraud statutes when AI is used to commit computer crimes, and directs Treasury to stand up an AI cybersecurity clearinghouse for identifying and remediating software vulnerabilities. For businesses that buy AI rather than build it, the practical consequence is timing: a voluntary 30-day pre-release review, applied to the models that matter most, inserts a month between "the lab is finished" and "you can use it."
DeepSeek drops the floor under agentic coding again
DeepSeek moved its V4-Flash API out of preview and into public beta yesterday with the 0731 build, holding the price where it was: $0.14 per million input tokens on a cache miss, $0.0028 on a cache hit, and $0.28 per million output tokens, with a concurrency limit of 2,500. The model keeps the architecture and parameter count of the April preview — 284 billion total parameters, 13 billion activated per token, a one-million-token context window — and ships under an MIT license with ungated weights on Hugging Face. This is a post-training upgrade, not a new design.
The gains are concentrated exactly where agents live. On Terminal-Bench 2.1, the 0731 build scores 82.7, up from 61.8 for the preview and 72.1 for the larger V4-Pro preview, putting an open-weight model within a few points of the top proprietary coding scores. Cybergym lands at 76.7 and Toolathlon-Verified at 70.3, but two harder agentic evaluations stay low — Agents' Last Exam at 25.2 and AutomationBench Public at 25.1 — a reminder that long-horizon autonomy is unsolved even as tool-calling and terminal work get cheap.
DeepSeek beat its own larger preview model on every agentic benchmark it published at about a third of the output price. A day after OpenAI's 80 percent cut on its cheapest frontier model, the direction is unambiguous: what you spend now depends on how many tokens your agents burn, not what each one costs.
Amazon raises its build to $220 billion and still runs out of room
On its second-quarter earnings call, Amazon raised 2026 capital expenditure guidance from roughly $200 billion to $220 billion, with CEO Andy Jassy naming memory costs as the driver. AWS posted $42.2 billion in quarterly revenue, up 37 percent and its fastest growth in eighteen quarters, with operating income of $16.6 billion, up 64 percent. Contracted cloud backlog reached $496 billion, a $132 billion increase in one quarter. Jassy told investors the company will not have enough capacity to meet all its 2026 demand, and expects the same in 2027.
The constraint is spilling into unrelated markets. Apple reported quarterly revenue of $109.42 billion, up 16 percent, with iPhone revenue of $54.25 billion, up 22 percent — and flagged component availability as a risk, because AI data center construction is competing for the same parts. Sony, reporting 1.38 million PS5 units against 95.3 million cumulative, told investors it has secured memory supply through the fiscal year. When a console maker treats DRAM procurement as headline news, the AI buildout has stopped being a technology story and become a supply chain one.
One percent of résumés are now carrying hidden instructions
A research collaboration among Duke University, hireEZ, Arizona State, UC Berkeley, and UNC Chapel Hill examined 200,000 real résumés submitted to the hireEZ platform between July 2019 and December 2025 and found roughly one percent contained hidden prompt injection attacks: instructions embedded invisibly in the PDF, in near-invisible type or blended into the background, telling an automated screener to mark the candidate as qualified. Incidence rose sevenfold between July 2024 and November 2025 across multiple industries. Duke's Neil Gong said what surprised the team was "how quickly the tactic is spreading."
The researchers declined on ethical grounds to test whether the injections changed hiring outcomes, so the success rate is unknown — which is the wrong thing to take comfort in. Any business using an AI screening tool, an inbox summarizer, or an agent that reads inbound documents is in the same position. Résumés are simply where someone finally counted.
AI assistants know your brand and still won't name it
Victorious published its Q2 2026 search report on July 29, testing 175 brands across legal, healthcare, SaaS, financial services, and ecommerce on eight AI platforms. Asked directly, the assistants described 96 percent of brands accurately. Asked the questions buyers actually ask — comparison and category-research prompts — 89 percent never appeared at all. Recognition and recommendation are almost entirely separate systems.
The mechanism is citation, not memory. Across 49,391 citations generated by category-research prompts, 99.99 percent pointed to third-party sites; only four of 150 brands earned a citation to their own website. What correlated with getting mentioned was referring domains (0.49) and third-party web mentions (0.45) — signals a company does not control directly. Brands with fewer than 2,000 indexed mentions surfaced three percent of the time. Platform behavior varied: Google AI Mode, Gemini, ChatGPT, AI Overviews, and Copilot each cleared 83 percent recognition accuracy, while Perplexity fell below 55 percent for SaaS and ecommerce.
Identity becomes the agent security perimeter
Okta agreed on July 30 to acquire Permiso Security for just under $200 million in an almost all-cash deal expected to close in the third quarter of its fiscal 2027. Permiso does identity threat detection across human, non-human, and agentic identities in multi-cloud environments — it watches what an identity does after login rather than only verifying it at the door. That is the correct shape for the agent problem: an agent authenticates once and then acts thousands of times, and the interesting failures all happen after authentication.
The same week produced a reminder of how far a single credential can reach. Wiz Research disclosed CosmosEscape, a chain of flaws in Azure Cosmos DB's Gremlin query engine. Queries were translated into .NET code with restrictions that did not account for .NET reflection, letting researchers build arbitrary code execution from queries against their own database — and extract the Cosmos Master Key, a platform-wide secret that could retrieve the primary key of any Cosmos DB account. Wiz reported it in November 2025; Microsoft hotfixed the entry point within 48 hours and completed a full architectural fix across all regions in July 2026, eliminating the master key from the design. Microsoft says it found no customer impact.
A German court draws the first hard line on generative music
The Munich Regional Court's 42nd Civil Chamber, under presiding judge Elke Schwager, ruled yesterday that Suno infringed the copyrights of songwriters represented by GEMA, the German collecting society. The court held that storing works inside the model violates the reproduction right and that serving outputs violates the making-available right — meaning licenses are required both for training on the repertoire and for commercially exploiting music generated from it. Suno was ordered to disclose revenue tied to the infringing activity and pay damages, with the amount still to be determined. The judgment is not final and can be appealed.
Quick Takes
Reddit posted $805 million in second-quarter revenue, up 61 percent, and the stock still fell more than 9 percent premarket on "choppy" search engine referrals and slowing US user growth — a distribution-risk lesson for any business whose traffic depends on a platform it does not own.
Google added an AI image tool to Google Earth built on its Nano Banana 2 model, letting users transform real locations from a text prompt, with SynthID watermarking on generated images.
Microsoft set an interim Xbox target of 500 million daily players by 2030 across PCs, mobile, cloud streaming, smart TVs, and subscriptions.
Thinking Machines co-founder Lilian Weng stepped down citing health issues and has rejoined OpenAI to lead a team focused on accelerating internal research.
The AI wearable Friend returned as Friend 2.0 with a built-in speaker and consistent voice at $249, up from $99. Hint, a home-management assistant co-founded by Martha Stewart, launched to track maintenance and organize property documents.
Amazon's custom-chip line reaching a $25 billion run rate makes Trainium and Inferentia a material alternative to Nvidia for inference, not a science project.
A separate 60-day item in the same executive order required the Office of Personnel Management to expand federal Tech Force cybersecurity hiring pathways.
What This Means for Your Business
Treat inbound documents as hostile input. The résumé finding is the visible edge of a general problem: if you have pointed a model at anything a stranger can send you — applications, invoices, support tickets, RFP responses, contract redlines — you have connected untrusted text to a system that acts on text. The fix is architectural, not vigilance. Strip formatting and hidden layers before a model sees a file, render PDFs to plain text you can inspect, keep the screening model advisory rather than decisive, and require a human to see any document the model flags as unusually eager to be advanced. If your applicant tracking vendor cannot tell you what it does about this, that is your answer.
Rebuild your marketing measurement around citations, not rankings. The Victorious data says the thing your business most needs to hear: being known is not the same as being recommended, and your own website is nearly irrelevant to whether an assistant names you. Ninety-nine point nine nine percent of citations in buyer-research answers went somewhere else. That reallocates budget. Third-party presence — industry directories, comparison sites, review platforms, trade publications, podcast and newsletter mentions, partner case studies — is now the acquisition channel, and the 2,000-mention threshold gives you a concrete target rather than a vibe. Run your five most important buying questions through ChatGPT, Gemini, Perplexity, and Copilot this week and write down who gets named. That list is your competitive set, whether or not it matches the one in your deck.
Rewrite your AI budget around volume, not unit price. Between OpenAI's cuts and DeepSeek's 0731 release, a capable agentic coding model now costs $0.28 per million output tokens with open weights and an MIT license. At that level the question is no longer "can we afford to run this on every ticket" but "what stops an agent from running ten thousand times when it should run ten." Put a hard spend cap and per-task token ceiling on every agent before you widen its scope, log token consumption per workflow the way you log spend per vendor, and pick your model per task rather than standardizing on one. Cheap open weights also make on-premises or private-cloud deployment realistic for regulated work that could not leave your network at last quarter's prices.
Plan hardware purchases early and identity governance now. Amazon's $220 billion budget, Apple's component warning, and Sony's memory procurement all point the same way: parts that go into servers, laptops, and phones are getting scarcer and more expensive through 2027. If you have a refresh cycle, a new office, or a storage expansion in the next eighteen months, order sooner and lock pricing rather than waiting for a better deal that is not coming. On the security side, Okta's purchase of Permiso reflects where the risk has moved. Inventory every non-human identity in your stack — service accounts, API keys, agent credentials, integration tokens — give each one the narrowest scope that works, set expirations, and monitor what those identities do after they authenticate. CosmosEscape is the reminder that a single over-privileged key can be the whole blast radius.
Finally, watch the frontier framework without over-reacting to it. Nothing binds your business today, and the voluntary review applies to a handful of labs. But if a 30-day pre-release government window becomes routine for the most capable models, your vendors' release cadence slows by roughly a month, and any roadmap that assumes a new frontier model lands the day it is announced needs a buffer. Meanwhile the Suno ruling puts real teeth into training-data provenance in the EU: if you use generative tools for music, images, or copy in European markets, ask your vendors what they license and get the indemnity in writing.