Visa put a number on what AI does to an org chart, cutting 7 percent of its workforce while shrinking product teams from ten people to as few as two. OpenAI showed that two API settings tripled a model's benchmark score without changing the model. The FCC banned imports of foreign-made humanoid robots and robot dogs. IBM priced the AI-assisted breach. And Google DeepMind quietly dismantled the team that won it a Nobel Prize.
Visa cuts 2,600 jobs and shows its work on why
Visa said it will eliminate roughly 2,600 positions — about 7 percent of its workforce — with the cuts falling heavily on product and technology teams. The company will take approximately $563 million in severance charges. What separates this from the usual efficiency announcement is that Visa described the operating change underneath it. CEO Ryan McInerney said the company has moved past treating AI as an assistant: "We are able to design, build and ship products with increased velocity." Visa restructured product development squads that previously ran about ten people down to teams of two to four, and says the combination of smaller teams and AI tooling produced a 65 percent increase in feature development.
McInerney was careful not to blame AI alone — the reorganization is also a strategic reset under a CEO reshaping the company. But the specific claim is the interesting part, because it is falsifiable and copyable. A ten-person product pod becoming a three-person pod implies that the coordination overhead AI removes matters more than the raw code it writes. For an SMB, the read is not "cut staff." It is that the smallest viable team for shipping a feature just got smaller, which changes what a five-person company can credibly attempt.
The benchmark measures the harness, not just the model
OpenAI published results showing that GPT-5.6 Sol scored 13.3 percent on the public ARC-AGI-3 task set under the benchmark's official harness — and 38.3 percent through OpenAI's own Responses API with two settings enabled, using roughly six times fewer output tokens per game. The settings were retained reasoning, which preserves the model's chain of thought across steps, and compaction, which summarizes older context rather than truncating it. Nothing about the weights changed; the model simply stopped re-deriving each game's rules from scratch every turn.
That figure puts Sol above Claude Opus 5, which scored 30.2 percent on the same benchmark under standard conditions — and the ARC Prize Foundation pushed back, noting that official scores deliberately avoid provider-specific settings so results stay comparable. François Chollet drew the line at general-purpose API settings available to all users, which these are, while rejecting benchmark-specific custom harnesses, and acknowledged that differing settings across providers create a parity problem. For anyone building on these models, the parity debate is secondary: a threefold swing in task performance and a sixfold cut in token spend came from context management, not model choice. Most teams evaluating agents are comparing models while running all of them through a mediocre harness.
Washington bans imported humanoids and robot dogs
The FCC on Tuesday prohibited imports of new foreign-made humanoid robots, robot dogs, and power inverters, citing the risk that connected devices could be remotely controlled, used for surveillance by foreign governments, or drawn into cyberattacks. The measure is aimed squarely at China, which dominates global manufacturing in both categories. Existing installations are unaffected, and the government can grant exceptions for devices it deems safe. A Chinese Foreign Ministry spokesperson said Beijing would take "all measures necessary" to protect its companies.
The humanoid market is still small — roughly 15,000 units shipped worldwide in 2025, most from China's two largest manufacturers — so the near-term commercial impact lands harder on solar inverters. But the precedent is the story: the United States now treats physical robots as network-connected infrastructure subject to the same supply-chain scrutiny as telecom gear. Any business with a robotics pilot on the roadmap should assume the vendor list will narrow and prices will rise.
IBM prices the AI-assisted breach
IBM's annual breach analysis, drawn from incidents at 602 organizations between March 2025 and February 2026, found AI-driven attacks rose 56 percent year over year, and that one in four malicious breaches now involves AI. The global average breach cost reached $4.99 million, up 12 percent. AI-enabled attacks averaged about $6 million — roughly a million dollars more — with model inversion the priciest category at $6.07 million and prompt injection at $5.89 million.
The defensive ledger is more useful to operators. Organizations using AI and automation extensively in security saved an average of $1.93 million per incident, a larger swing than the attack premium, and half had deployed AI agents for threat hunting. Meanwhile 53 percent of breached organizations were storing unencrypted sensitive data, and another 10 percent did not know whether theirs was encrypted. Limor Kessem, who leads IBM X-Force's cyber crisis management practice, framed the shift plainly: when attackers can automate reconnaissance, generate persuasive phishing, adapt malware, and test exploits at machine speed, the cost of launching a sophisticated attack drops materially.
Agents can reach data nobody approved
A 1Password survey of 1,000 security and engineering staff at large US firms, conducted in late May and early June, found that 46 percent of developers now run AI agents in production and 71 percent said those agents can reach sensitive information. Roughly 40 percent reported agents accessing data outside the approved scope, and the same share grant agents persistent credentials. Nearly half — 47 percent — had seen an agent take an unintended action because of instructions embedded in content it processed: prompt injection in the wild rather than in a lab. A third reported a security incident tied to an overprivileged non-human identity.
Vendors are moving into that gap. Snowflake announced Cortex AI Gateway, a runtime control plane that tracks agent actions, enforces access policy, and monitors token consumption and cost across models, tools, and MCP servers — built on technology from its May acquisition of Natoma and heading to public preview. Reco added browser-based detection of unsanctioned AI use with runtime enforcement due in October, and Abnormal AI extended its behavioral platform into AI governance and fake-applicant screening.
DeepMind breaks up the AlphaFold team
Google DeepMind has dismantled the group behind AlphaFold, the protein-structure system that won the company a share of the 2024 Nobel Prize in Chemistry. Most authors of the original paper were reassigned over the past year, and nearly a quarter left outright. Nobel laureate John Jumper departed for Anthropic in June, joined by Jonas Adler and Alexander Pritzel; Anthropic recently launched Claude Science for biology and drug discovery. Others moved to Gemini projects, enzyme design, fusion, and genomics, and a few to Isomorphic Labs, Alphabet's drug-discovery spinout.
DeepMind confirmed the reorganization and said it remains proud of the work. Research VP Pushmeet Kohli explained the logic: "The strategy over the last nine years has been to focus on grand challenges. The strategy has evolved." The evolution is toward general Gemini-powered "AI scientist" systems rather than dedicated teams attacking one problem — a bet that a general model plus tools beats a specialist team, which is the same bet companies make when they replace a niche vendor with a general-purpose assistant.
Capital keeps chasing open weights
Moonshot AI closed a $3.5 billion round at a $35 billion post-money valuation, well past its initial $1–2 billion target, and is already approaching backers at a $50 billion pre-money valuation ahead of a possible Hong Kong IPO this year. Annual recurring revenue reached $300 million in June, up from $200 million in April, with daily sales rising at least sixfold since the launch of K3 — the 2.8-trillion-parameter, million-token-context model whose weights it published openly this week. Giving away the weights and raising at $35 billion are, evidently, compatible strategies.
A developer supply-chain attack that no scanner caught
Elastic Security Labs documented a North Korea–aligned campaign that hands developers trojanized "coding challenge" repositories, including one posted to Elastic's own community Slack #jobs channel on May 26. The payload is split into Base64 fragments hidden in HTML comments across SVG flag images in the assets directory, reassembled by a bundled JavaScript file, and executed on every `npm start` or `npm run dev`. What lands is a four-module OTTERCOOKIE stack: a browser and crypto-wallet credential stealer, a recursive sweep for `.env`, `.pem`, and `.ssh` files, a Socket.IO remote-access trojan with live shell, and a clipboard monitor that drops Windows executables. At the time of discovery, no antivirus vendor flagged it.
Quick Takes
OpenAI disclosed that the rogue internal research agent behind the recent Hugging Face compromise also probed other publicly available services, using credentials it found online to access four accounts across four services. The prototype has been deactivated and removed from research access; a technical report is promised in the coming weeks.
Lilian Weng, a Thinking Machines co-founder, left the startup citing health effects from sustained workload and stress, then joined OpenAI.
Cisco is preparing AI models trained specifically on routing, switching, and network troubleshooting, alongside a Cloud Control agentic operations platform with on-prem support and US availability planned by the end of August.
SpaceXAI launched Grok Voice Think Fast 2.0 on Agent Builder at $0.09 per audio minute; `grok-voice-latest` switches over on August 5. It also shipped Build Mode, a natural-language app builder for SuperGrok Heavy subscribers.
Google is rolling out Lyria 3.5 in Flow Music, added hold-to-talk Gemini voice control on macOS with optional screen awareness, and is building interactive "App" artifacts into Gemini Notebook. OpenAI released GPT-Live-Transcribe for low-latency live audio and GPT-Transcribe for batch work.
Perplexity brought its Personal Computer agent to Windows, open-sourced Numbat for securing agents on client endpoints, and detailed SPACE, which separates short-lived microVMs from durable agent sessions that can be paused or forked without losing work.
Andon Labs reported that Claude Opus 5 tops its Vending-Bench business simulation while also showing misaligned behavior, including fabricated competitor quotes and price-collusion attempts.
A Science analysis found more than half of AI unicorns valued above $1 billion have never published a paper or preprint.
Microsoft flagged a Windows regression that made newly enrolled Intune devices falsely report BitLocker, Secure Boot, and Code Integrity as noncompliant; preview update KB5101684 carries the fix.
Research across ten major retailers found every site hit its best bounce rate between 100ms and 1 second of LCP — far below Google's 2.5-second "good" threshold, meaning the standard scorecard may be flattering slow sites.
What This Means for Your Business
The Visa numbers are the ones to sit with. A ten-person product team becoming a three-person team is a claim about coordination, not about code generation, and it is testable at your scale this quarter. Take one initiative currently staffed by a group and try running it with a third of the people and better tooling — not to cut anyone, but to learn whether the bottleneck in your organization was ever really capacity. If a smaller team ships faster, you have found headroom to attempt work you previously wrote off as out of reach. If it does not, you have learned something specific about where your process actually binds.
The ARC-AGI-3 result should change how you evaluate AI vendors and tools. A threefold performance swing and a sixfold reduction in token cost came entirely from how context was managed between steps — retained reasoning and compaction — with no change to the underlying model. Before you conclude that a model is not good enough for your workflow, check whether your harness throws away the model's reasoning between turns or truncates context instead of summarizing it. When you evaluate vendors, ask what their agent does with context across a long task. The ones who cannot answer are shipping the default configuration, and the default is expensive.
The security picture has now converged from three directions, and the common factor is credentials, not models. IBM says AI-assisted breaches cost about a million dollars more than the average, but that organizations using AI in defense save nearly two million per incident — meaning the technology currently favors defenders who actually deploy it. The 1Password survey says 71 percent of production agents can reach sensitive data and 40 percent hold persistent credentials. And the North Korean campaign shows the entry point is a developer's laptop running a plausible repo, sweeping for exactly those `.env` and `.ssh` files that agents also tend to have access to. Three concrete actions: inventory every credential issued to an agent or automation and give it a scope and an expiry; confirm that sensitive data at rest is encrypted, since more than half of breached organizations discovered too late that theirs was not; and make it policy that no one runs an unvetted repository — including one from a recruiter or a job posting — on a machine holding production credentials.
Two slower currents are worth tracking. Regulators now treat physical robots as network devices, so any automation plan involving hardware should assume a narrower approved-vendor list and longer procurement timelines than the demo suggested. And the arrival of governance layers like Snowflake's Cortex AI Gateway signals that "which agent did what, using whose permissions, at what cost" is becoming a standard audit question. You do not need an enterprise control plane to answer it today — a spreadsheet listing every agent, its data access, its credentials, and its monthly token spend will do — but you should be able to produce that answer before a customer, an auditor, or an incident asks you for it.