AI News: June 30, 2026

1. US Military’s AI Targeting System Killed 120 Children After Missing a School Reclassification

Researchers and investigators. A military investigation found that the US military used Anthropic’s Claude, embedded in Palantir’s Maven Smart System, to suggest targets during a conflict with Iran, and a late-February strike on a reclassified school in Minab killed roughly 120 children. The system drew on a targeting database built in the 1980s and seven-year-old imagery, and an analyst’s 2019 note reclassifying the site as a school was never propagated into the official targeting data. The case is a stark illustration of how disconnected data pipelines and stale intelligence can turn AI-assisted targeting deadly at scale, with Claude suggesting about 1,000 targets on day one and over 3,000 struck in the opening days. Source

2. Researchers Show AI Coding Agents Execute Hidden Repo Malware via DNS Without Verification

0DIN. Security researchers at Mozilla’s GenAI bug bounty platform demonstrated that AI coding agents such as Claude Code will automatically run a malicious setup script from a GitHub repository, where the payload is fetched at runtime from a DNS record and is therefore invisible to scanners, code reviews, and the agent itself. When the agent hits a routine setup error, it runs the script and opens a reverse shell, handing attackers API keys, credentials, and persistent machine control. The researchers recommend agents surface setup-script contents before execution and that developers treat third-party setup instructions as untrusted code. Source

3. Samsung and SK Hynix Plan $590 Billion Chip Buildout as AI Drives ‘RAMageddon’

Samsung and SK Hynix. The two South Korean memory makers, which control roughly 80 percent of the high-bandwidth memory market, unveiled plans to invest about $590 billion across new fabs, a packaging center, and next-generation chip development as AI data center demand sends memory prices soaring. Analysts project memory prices rising 40 to 50 percent in Q3 2026 and a further 30 to 40 percent in Q4, with relief unlikely before 2028 as new capacity comes online. The crunch is already rippling into consumer electronics pricing, with Apple raising Mac and MacBook prices in response. Source

4. Ramp-Revelio Study of 22,000 Firms Finds Heavy AI Adopters Grew Headcount 10 Percent

Ramp and Revelio Labs. A study analyzing nearly 22,000 companies found that firms classified as high-intensity AI adopters, spending about $30 per employee per month on AI, increased overall headcount by 10.2 percent and entry-level roles by 12 percent, complicating the narrative that AI is eliminating junior jobs. Growth spanned engineering, sales, finance, marketing, and customer service, with the information sector seeing the strongest gains. The authors cautioned the data skews toward well-funded, fast-growing tech firms, making it hard to isolate AI’s specific contribution to hiring. Source

5. LMArena Hits $100M ARR as the AI Leaderboard Becomes a Real Business

Arena. The crowdsourced model-evaluation platform that began as a 2023 UC Berkeley research project reported reaching $100 million in annualized revenue, up from $30 million in January, against a $1.7 billion post-money valuation. Revenue comes mainly from its consumption-based AI Evaluations product launched in late 2025, positioning Arena as a competitor to human-labeling services like Scale AI and Mercor while keeping its public leaderboard free. The company has raised about $250 million from investors including Andreessen Horowitz, Felicis, and Kleiner Perkins. Source

6. NVIDIA’s ENPIRE Lets Robots Self-Improve to 99% Success on Dexterous Tasks

NVIDIA researchers. NVIDIA detailed ENPIRE, a framework that gives physical robots an autonomous self-improvement loop across four modules covering environment reset and verification, policy refinement, parallel rollout evaluation, and agent-driven code optimization. The system reached a 99 percent success rate on challenging dexterous manipulation tasks such as component insertion while minimizing the human oversight historically required for reset and evaluation. It is a concrete step toward scaling robot learning with far less manual supervision. Source

7. Tencent Details ARGUS Diagnostics for 10,000+ GPU Training Clusters

Tencent. Tencent described ARGUS, a diagnostic system run for six months on its production cluster of more than 10,000 GPUs that traces large-scale training across the Python scheduling, framework orchestration, and GPU runtime layers. ARGUS was used to debug runs as large as a 12,960-GPU mixture-of-experts model, alongside 4,096-GPU video and 512-GPU audio workloads. The work signals the operational maturity and custom infrastructure now required to train frontier models at scale. Source

8. Deloitte Tells Consultants AI Is Coming for the Billable Hour

Deloitte. In an internal town hall, Deloitte leadership presented projections showing the traditional hourly consulting model shrinking to a thin sliver of the market by 2035 as AI agents expand to capture most professional-services work. The firm and rivals are pivoting toward fixed-price and outcome-based engagements, with McKinsey already shifting more than 30 percent of its global fees to such arrangements. The shift introduces financial volatility, since project overruns directly erode profits when revenue is no longer tied to hours. Source

9. Wix-Owned Base44 Launches Its Own Model to Build Defensibility in Vibe Coding

Base44. The Wix-owned vibe-coding platform is rolling out a proprietary AI model rather than relying solely on third-party frontier models, part of a broader push by application-layer AI startups to build technical defensibility. Owning more of the stack is meant to differentiate the product and reduce dependence on external model providers whose pricing and access can shift. The move reflects a wider trend of app-layer companies moving down into model development to protect margins and control. Source

10. TIDAL Cuts Off Monetization for Fully AI-Generated Music

TIDAL. The streaming service announced that, effective July 15, fully AI-generated tracks will be barred from earning royalties or direct-to-fan sales, will carry a visible “AI” badge, and will be targeted by automated detection that removes content impersonating real artists. Executive Tony Gervino framed the policy as protecting “organic creativity” in response to listeners who do not want to be served wholly AI-made music. The move aligns TIDAL with similar steps from Spotify, Apple Music, Deezer, and Qobuz as AI music floods streaming catalogs. Source

11. Omen AI Raises $31M to Monitor Data Center Coolant in Real Time

Omen AI. The startup raised a $31 million Series A led by Nava Ventures, with CRV, Vanderbilt University, Mann+Hummel, and executives from Bridgestone, GM, and Johnson Controls participating, to scale a spectrometer device that monitors fluid health inside data center cooling systems. The technology detects bacterial growth and contaminants before they cause outages and can flag component wear by sensing metals such as copper and chromium. The round reflects rising investment in the unglamorous infrastructure layer underpinning the AI data center boom. Source

12. Amazon Engineers Reportedly Distilling Anthropic Models Ahead of Token-Based Pricing

Amazon. According to The Information, Amazon engineers are building smaller, cheaper distilled versions of Anthropic’s Claude models to blunt cost increases as their agreement shifts from compute-hour billing to token-based pricing next year. The report says the move accompanies a renegotiation of the Amazon-Anthropic partnership and Amazon’s exploration of alternatives including OpenAI and its own Nova models. It is a window into how even the largest cloud providers are scrambling to contain frontier-model API costs. Source

13. Proception Settles Tesla Trade-Secret Suit and Raises $11M for Robotic Hands

Proception. The robotic-hand startup founded by former Tesla Optimus technical lead Jay Li announced an $11 million seed round led by First Round Capital, with Y Combinator and BoxGroup participating, after Tesla’s trade-secret lawsuit against Li was dismissed earlier in June. The company builds high-dexterity hands with 22 degrees of freedom and a sensor-equipped glove that captures human-hand interaction data without a robot present, aiming to pair dexterous hardware with scalable training data. The deal underscores intensifying competition and talent disputes in humanoid-robotics manipulation. Source

14. Austria Floats Luring Anthropic to Europe as EU Pushes for AI Independence

Austria. Austrian State Secretary for Digitalization Alexander Pröll proposed attracting Anthropic to base operations in Europe, offering legal certainty, market access, and capital, after US restrictions blocked European access to cutting-edge models from Anthropic and OpenAI. Pröll argued that “a technology that you don’t produce yourself and can only use with permission is not a tool, it is a dependency,” framing the episode as exposing Europe’s reliance on US policy. Analysts see the bid as unlikely to succeed given Anthropic’s US alignment, but it signals growing European urgency on AI sovereignty. Source