the Fable 5 export-control standoff dominating, 42 state attorneys general subpoenaing OpenAI, and SpaceX buying Cursor
Apertus launched as an open foundation model designed for sovereign deployment: permissive weights, no US export restrictions, and a training process designed to be auditable by national governments and enterprises with data-sovereignty requirements.
A widely shared essay argued that developers should reject AI-generated code even when it passes tests if the author cannot explain its intent, trace its behaviour or maintain it six months later, framing "it works" as an insufficient bar for merging agentic output.
TechCrunch surveyed the AI features in the iOS 27 developer beta beyond the headline Siri changes: on-device summarisation across apps, AI photo editing without a server round-trip, and smarter autocomplete using on-device Gemini inference via Core AI.
A week into the ban, Politico reported the White House and Anthropic talks have moved from "fix the jailbreak" to designing a formal framework for grading the severity of AI security flaws, an acknowledgement that no model can be made fully hack-proof. Export-control experts questioned the legal basis for blocking foreign nationals from a cloud login (former Commerce official Kevin Wolf said he did not know what the authority was, and UC Berkeley's Andrew Reddie called jailbreak-immunity-as-standard unsustainable). The New York Post reported Anthropic has pledged closer White House cooperation and faster security response, and its international managing director said in Seoul he was "very confident" Fable 5 and Mythos would be available again "within days". Continues the standoff logged 15-19 June.
Commerce Secretary Howard Lutnick told ASML the US government believes one of its EUV ultraviolet lithography machines may have made its way into China, with senior officials saying ASML was "not acting in good faith". The report, via Bloomberg, is still developing and lands amid the wider US-China struggle over AI compute.
Senator Bernie Sanders unveiled legislation to create a $7 trillion sovereign wealth fund financed by a one-time 50% tax on the equity of any company with more than $200 million in annual AI revenue, a list reaching well beyond OpenAI and Anthropic into public companies and startups. A presidentially appointed commission would hold voting shares, appoint board members and block corporate decisions (de facto nationalisation) while paying every American over $1,000 a year. JD Vance said the president is sympathetic to the sovereign-wealth idea in principle but parts ways on the redistribution mechanism.
Accenture reported a 2% drop in bookings and missed revenue, sending its stock down 18% on Thursday to a near-decade low, with the shares now halved this year. The company blamed a $400 million hit to its Middle East business from the Iran war, but critics tied the miss to the AI-transformation story, with rivals arguing that real AI implementation needs domain expertise large consultancies lack.
OpenAI's Codex shipped Record and Replay: demonstrate a recurring task once and Codex encodes the demo as an inspectable, editable skill. Microsoft's Nicolas Bustamante flagged its value for legacy systems, noting you can record actions on a computer and have the AI do the work end to end, which matters for old software with no APIs.
Microsoft shipped a serverless agents runtime for Azure Functions in public preview at Build 2026, with agents defined in `.agent.md` markdown files, extending its serverless platform into hosting AI agents.
GitLab 19.0 extends agentic AI beyond code generation into securing credentials and secrets, reviewing and merging changes, and scanning dependencies for supply-chain security, pushing agents deeper into the delivery pipeline.
CircleCI launched Chunk Sidecars, a capability that brings CI-style validation directly into an AI coding agent's inner development loop so that agent-written changes are checked as they are produced rather than only at merge.
A Windows Developer Blog post, "Windows platform security for AI agents", set out Microsoft's pitch for Windows as a trustworthy operating layer for running AI agents safely, part of the wider race to harden the systems agents act within.
Cloudflare introduced temporary, task-scoped accounts for AI agents: short-lived credentials issued per session that expire automatically, preventing an agent's access from persisting beyond its current task and reducing the blast radius of a compromised or misbehaving agent.
The Valkey open-source Redis fork replaced manual backporting of upstream bug fixes with AI agents that detect relevant patches, apply them to older release branches, run the test suite, and open a pull request, cutting maintainer overhead on a task that previously consumed several engineer-days per release.
Pinterest is moving past the keyword search bar with Ask Pinterest, a conversational shopping experience, alongside a new Business Assistant and an MCP layer that pipes the platform's taste and intent signals straight into agents.
Sam Altman has signed Noam Shazeer, a lead author of the 2017 "Attention Is All You Need" transformer paper, away from Google, less than two years after Google paid around $2.7B to bring him back through the 2024 Character.AI deal. Altman called it a hire he had wanted since OpenAI's founding, and the exit, with Gemini 3.5 Pro slipping past its expected June window, has fed speculation about Gemini's direction.
The G7 in France, the most AI-heavy yet, produced no concession on the Fable 5 and Mythos 5 ban: Dario Amodei and Demis Hassabis pressed for a US-led coalition and structured model access, but Washington gave nothing and Trump's only public line was "going fine". Reporting filled in the backdrop, including a Lutnick letter warning Anthropic against distributing the models to "foreign persons", leaked internal messages from staff who feel "unfairly targeted", and Wired naming SK Telecom as the Korean carrier whose Mythos access triggered the revocation. Macron sought a workaround and Keir Starmer was refused a carve-out for British users. Continues the standoff logged 15-18 June.
A week into the Fable 5 shutdown, builders are stitching together open weights and model routing to approximate Fable performance at lower cost, with the AI Daily Brief surveying Kimi K2.7-Code, GLM-5.2, Microsoft serving DeepSeek, Cursor's Composer 2.5, OpenRouter Fusion and legal-AI firm Harvey's routing. Moonshot's Kimi K2.7-Code claims a 22% gain on its CodeBench and 30% fewer reasoning tokens, and one landing-page test found it about 16 times cheaper than Fable 5, though practitioners say the benchmarks oversell real-world gains.
Latent Space reported GLM-5.2 passing the community "vibe check" and edging GPT on frontend coding, with Z.ai forecasting an open-weight model approaching Fable-class capability by December, a marker of how fast the open frontier is closing on the shut-down US models. Extends the GLM-5.2 thread logged 15-17 June.
OpenAI is sunsetting Pulse, the personalised daily briefing it launched last year, and rolling its faster, more reliable scheduled-tasks feature out to all paid tiers including the cut-price Go plan, the latest in a run of feature retirements that has prompted some subscribers to question the value of higher tiers.
Cybersecurity firm Chainguard announced Athena, an industry coalition using AI to find and fix vulnerabilities in widely used open-source software before attackers can exploit them, focused on the libraries, containers and components that underpin browsers, data centres, phones and payment systems.
OpenAI is readying GPT-5.6, reportedly with Mini and Pro variants, for release as soon as next week, adding a larger 1.5M-token context window, stronger long-horizon coding and faster Codex responses, with pricing pitched to undercut Anthropic while Claude Fable 5 stays unavailable in the US.
Anthropic added "artifacts" to Claude Code, turning a work session into a live, shareable visual page (for things such as PR walkthroughs and system explainers) that refreshes automatically and carries version history and privacy controls, available in beta for Claude Team and Enterprise.
Perplexity introduced Brain, a memory system that builds a persistent context graph across tasks, projects, decisions, files and sources so agents start with relevant context rather than from scratch, linking each memory to its original source and claiming better answer accuracy and lower task costs through reuse.
Amazon's AWS is in talks to sell its Trainium AI chips to other companies for use in their own data centres, a more direct challenge to Nvidia, though with its current capacity already sold out the move could leave existing customers waiting unless it can make more.
Google is renting computing power from thousands of its own TPUs at a Western New York data centre to Anthropic, using Nvidia's playbook to build a rival AI chip business; it spotted the commercial potential of its TPUs about two years ago and is now focused on improving their inference performance.
Meta's Yann LeCun dismissed Elon Musk's xAI as a failure unable to compete with OpenAI and Anthropic, arguing Musk struggles to hire top talent and must rent infrastructure to recover costs, and warned that AI labs must raise prices or cut costs to avoid a "big bubble explosion" given their high operating expenses.
A study in NEJM AI from Boston Children's Hospital, Harvard and OpenAI ran the o3 Deep Research reasoning model over 376 previously unsolved cases and, after expert review and clinical confirmation, helped establish diagnoses in 18 children, a 4.8% additional diagnostic yield, spanning neurodevelopmental, neuromuscular and other conditions. The model produced evidence-linked hypotheses for specialists to investigate rather than diagnosing on its own.
Mistral is expanding its Vibe product with a Code section for browser-based coding and an Apps area for building and sharing apps, pushing it beyond basic chat.
Emily Dalton Smith, the Meta executive leading its internal AI overhaul, is leaving after about two months, staying on to hand over to a replacement, the latest sign of turbulence in Meta's AI-driven reorganisation. Extends the Meta engineering reorg logged 17 June.
Norway announced a near-ban on AI tool use in state-funded primary schools, with the government citing evidence that children who outsource recall and reasoning to AI before developing foundational skills show measurable deficits; the policy takes effect in autumn 2026 and is one of the most restrictive AI-in-education policies in Europe.
A peer-reviewed Nature study reported early evidence that AI-assisted cognitive work leads to measurable skill atrophy: participants who used AI for recall and reasoning tasks over an extended period scored significantly worse on those tasks without AI than a control group.
The Financial Times reported that a growing number of large enterprises are reducing or capping AI tool spending as ROI cases prove harder to close than expected; several firms are consolidating onto a single AI provider to contain costs, and CFOs are requesting proof of measurable productivity gains before renewing contracts.
An AI engineer published a proposed partial decipherment of Linear A, the undeciphered Minoan writing system, using GPT-5 combined with linguistic analysis of known Minoan loanwords in Linear B; the claim of a 40-60% vocabulary match is disputed and awaiting peer review.
GitHub closed its free AI Model Playground to new signups, directing teams toward GitHub Copilot and the broader Codespaces environment; existing users can continue for now with no announced sunset date.
An O'Reilly-sourced Stack Overflow Blog essay argued that high-stakes AI agents should adopt a contract-based design: declaring capabilities and obligations up-front, accepting structured constraints from deployers, and producing auditable records of whether they met those contracts during execution.
Checkmarx's new static analysis engine is built around the assumption that LLM-generated code is the dominant input: it prioritises patterns typical of AI output (verbose boilerplate, latent permission escalations, inconsistent error handling) and integrates directly with the agent's review loop rather than running as a separate gate.
Pew Research's 2026 survey of more than 5,000 US adults found adoption and optimism moving in opposite directions: about half now use a chatbot and a quarter do so daily, up from roughly a third of the public in 2024, even as trust in AI continues to slide.
Anthropic shipped a major Claude Design update that lets users import design systems from a GitHub repo, design files or raw uploads and round-trip between design and code, repositioning the tool from a prototype toy to an enterprise platform while addressing the heavy token consumption that dogged the earlier version.
Vercel launched eve, an open-source agent framework with durable execution, sandboxed compute, approvals, subagents and evaluations built in, alongside Connect, which replaces long-lived provider tokens with short-lived, task-scoped credentials issued at runtime to cut the risk of shared agent access tokens.
Ars Technica reported that AI coding agents in Nvidia's GEAR lab can autonomously direct robot training, with tasks such as installing GPUs and cutting zip ties, and the lab plans to open-source its ENPIRE harness so anyone can host a self-directing robot lab.
A study highlighted by The Neuron argued that domain expertise rather than raw coding skill is the better predictor of who gets strong results from Claude Code, reinforcing the wider finding that judgement and context, not typing speed, are becoming the scarce inputs in agentic engineering.
Apple is preparing to raise product prices as surging memory and storage costs, driven by the AI infrastructure buildout consuming supply, work through its bill of materials; prices have quadrupled since last year and are expected to keep climbing into 2027.
Mobileye said it will enter the US robotaxi market in 2027 with a vertically integrated service built on its Moovit platform, starting with about 100 vehicles in an unnamed city and aiming to scale to roughly 17,000 within five years if the pilot succeeds.
OpenAI introduced LifeSciBench, an expert-judged benchmark that evaluates AI systems on end-to-end life-sciences work, including evidence analysis, experimental design, scientific reasoning and research communication, rather than isolated biology tasks.
Five days into the US shutdown of Fable 5 and Mythos 5, Anthropic sent a technical delegation (Tom Brown, Sarah Heck, Logan Graham, Nicholas Carlini) to Washington, but none of the officials said to have driven the ban attended. Commerce reportedly would let Fable return if the jailbreak were fixed, yet treats Mythos separately after Anthropic added around 50 firms to Project Glasswing, including a South Korean telecom suspected of Chinese-government ties. Bloomberg published Commerce Secretary Lutnick's letter threatening criminal and civil penalties, and AI chiefs flew to the G7 in France to seek a resolution. Continues the saga logged 11-17 June.
A Pentagon filing supporting the DOJ's move to dismiss the NAACP suit over xAI's Colossus II gas turbines confirmed Grok was used for targeting in recent Iran strikes, enabling "over 2,000 munitions to 2,000 distinct targets within 96 hours". The DOJ argued cutting xAI's power would threaten national, economic and energy security; Anthropic, Google and OpenAI models are also cleared for classified use, and Dario Amodei has said Claude was used in missile targeting. Extends the xAI data-centre suit logged 16 June.
Midjourney announced the Midjourney Scanner, a full-body ultrasonic CT system that founder David Holz called the "first new whole-body medical imaging modality in 50 years", using 358,000 ultrasonic elements across 40 ringed systems rather than MRI or X-ray. The Gen 1 prototype takes about 20 minutes per scan and shows no AI in the imagery yet, though reconstruction and downstream uses will rely on it. The self-funded company has begun FDA talks, plans a 25,000 sq ft "Midjourney Spa" of scanners in San Francisco by the end of 2027, and targets a fleet of 50,000 machines doing up to 1 billion scans a month.
AWS used its latest updates to pitch "continuous agentic DevOps", putting its Kiro coding agent "in your pocket" and adding Bedrock AgentCore features its AI chief said would prevent bad outcomes, while conceding that trust is the biggest barrier to AI adoption.
Britain's competition regulator ordered Google to make its search rankings less opaque under new rules covering organic rankings, AI Overviews and user-approved sharing of search data, part of its strategic-market-status oversight of the company.
The Register reported that only about half of the US data-centre capacity planned for 2026 is actually being built, another sign of AI infrastructure ambitions colliding with reality.
A new MCP specification extension adds structured OAuth-style authorization: agents must declare the permission scopes they need before a session begins, users or operators approve them explicitly, and access tokens are scoped to the session and expire at its end, addressing the top enterprise security objection to MCP deployment.
After SpaceX acquired Cursor, all three companies argued GitHub's current architecture is unsuited to agent-native development but disagree on what comes next: Cursor's Origin is a Git-compatible agent forge, GitLab is extending its CI/CD pipeline into an agentic SDLC, and Zed is betting on real-time collaborative editing as the new default developer surface.
Researchers published a compute-allocation optimization framework for coding agents that dynamically redistributes inference budget across planning, coding and verification stages; it outperforms Claude Code and Codex by 2.5x on shared benchmarks at the same compute budget.
Adobe embedded agentic AI workflows across Photoshop, Premiere Pro and After Effects: agents can now plan and execute multi-step production tasks (colour grading, audio cleanup, asset resizing) across apps autonomously, with the human reviewing a completed pipeline rather than each step individually.
A deep-dive on Z.ai's GLM-5.2, a 744B-parameter MoE with 40B active parameters, placed it just behind Opus 4.8 overall but ahead of every Opus (including 4.8) at frontend coding, and made it the first open-weight model past 80% on Terminal-Bench. Its headline systems contribution is IndexShare, a sparse-attention scheme that reuses one indexer across every four layers for about 2.9x lower per-token FLOPs at 1M context, plus a multi-token-prediction tweak lifting speculative-decoding acceptance by up to 20%, framing the release as much as an inference-efficiency package as a model update. VentureBeat separately reported GLM-5.2 beating GPT-5.5 on multiple long-horizon coding benchmarks at about one-sixth the cost. Extends the GLM-5.2 launch logged 15 June.
GitHub introduced the GitHub Copilot app, a desktop control centre for "agent-native" development that lets engineers launch and monitor several agent sessions in parallel from a "My Work" view wired into GitHub workflows, pitched at keeping developers in charge as agents take on more coding while taming the context switching and agent-code review that come with it.
At its Compile event, days after the SpaceX deal, Cursor teased a new frontier model (reported as Composer 3) said to exceed 1.5 trillion parameters, trained from scratch on more than 100,000 GPUs with 10-20x the compute of Composer 1, in the Opus and GPT-5.5 class and due within weeks. It also showed Cursor Origin, a Git-compatible "agent-native" GitHub rival built for fleets of agents cloning, branching and reviewing in parallel. Extends the SpaceX-Cursor acquisition logged 16 June.
DeepSeek raised more than $7.4B at a valuation above $50B, making it China's most valuable AI startup, with founder Liang Wenfeng putting in around $3B and a government-backed fund about $150M. The company says the capital will fund R&D and more computing infrastructure.
DX's quarterly analysis of more than 400 organisations found self-reported AI-authored code jumped to 51.9% in Q2, up from 27.4% in Q1, holding near 50% across company sizes. Over the past year the median pull request nearly doubled, from 44 to 72 lines, which Abi Noda's team flags as a growing review-and-quality risk as verbose AI output compounds technical and cognitive debt.
Android 17 began rolling out to Pixel devices with new agent capabilities centred on AppFunctions and Android MCP, letting apps expose orchestratable tools that on-device Gemini agents can discover and run, part of Google's pitch for the phone as an "intelligence system".
OpenAI described Deployment Simulation, a pre-release evaluation that replays real conversation contexts against candidate models to estimate how they will behave once deployed and to surface new safety issues before launch.
NVIDIA's Blackwell platform topped MLPerf Training 6.0 with the fastest training times and the largest-scale run at 8,192 GPUs, with its NVLink and NVFP4 features cited for efficiently routing mixture-of-experts models.
Genesis AI, backed by former Google CEO Eric Schmidt, unveiled Eno, a general-purpose robot it says can reason, adapt and own outcomes beyond predefined tasks, and is working with LG's services arm to deploy it to industrial customers by year end while raising its next round.
Reporting on Meta's engineering organisation described an AI morale crisis, arguing that a wave of changes since April has rapidly dismantled the "move fast" culture that defined the company as it leans harder on AI and restructures teams.
Apple is reportedly developing AirPods with built-in cameras that would feed Siri a view of what the wearer sees, targeted for a late-2027 launch alongside a foldable iPhone, as it builds out AI wearables.
AWS released an autonomous agent that continuously scans production codebases for deprecated libraries, outdated patterns and security debt, and opens pull requests to modernise them without a human trigger, integrating with CodeCatalyst and Bedrock AgentCore.
Ars Technica published an analysis arguing that models with advanced autonomous hacking capabilities will be widely available regardless of US export restrictions, because the capability emerges naturally from scale and top AI labs outside the US are rapidly closing the gap with no reciprocal controls.
Stack Overflow Blog described how an attacker compromised 20,000 Instagram accounts by asking a Meta AI agent to transfer data using natural-language requests that bypassed safety checks, framing the failure as an access-control gap: agents inherit their creator's permissions with no per-action authorization check.
More than 100 cybersecurity executives and researchers signed an open letter urging the US to lift its export ban on Anthropic's Fable 5, arguing it handcuffs defenders while attackers can pull the same flaw-finding capability from rival models (the letter notes OpenAI's Daybreak, GPT-5.5, Kimi 2.7, Opus and Sonnet all do it), with Alex Stamos saying the flagged "jailbreak" was really a proof-of-concept of a bug that defensive teams use to patch. The Register reported the feds reacted to a plain "fix this code" prompt rather than a jailbreak, Washington talks have not switched the models back on (they stay dark worldwide, with G7 allies blocked), and Brussels cast the curbs as proof Europe needs its own AI sovereignty. Continues the saga logged 11-15 June.
SpaceX is buying the AI code editor Cursor (Anysphere) in a roughly $60B all-stock deal days after its blockbuster IPO, folding it into Musk's frontier-model push, with a jointly trained model expected to land in both Cursor and Grok Build.
Salesforce acquired customer-support AI specialist Fin (formerly Intercom) for $3.6B, adding its support agents and about 30,000 customers to the Agentforce lineup; Fin claims its agents resolve roughly three-quarters of customer queries without human help.
Sakana released Marlin, its first commercial product, an autonomous research agent that turns a topic into a detailed strategy report and summary slides with no further human input and can run for up to eight hours in a single pass; it was beta-tested with around 300 industry experts. VentureBeat framed it as an "ultra-deep" research agent capable of 100-page reports in eight hours.
Meta added AI Mode to Facebook, a chatbot-style search powered by its Muse Spark model that answers questions from public Group posts, Reels and Marketplace content instead of returning links, alongside AI photo presets and auto-collages; Meta is reportedly lining up two paid AI tiers at $7.99 and $19.99 a month, undercutting ChatGPT and Gemini.
A dormant feature in the iOS 27 developer beta would let users replace Siri's AI brain with ChatGPT, Claude or Gemini from Settings, complete with a dedicated App Store section, though it was never announced at WWDC. OpenAI is reportedly weighing legal options, including a breach-of-contract notice over its existing Siri partnership, and the feature is blocked in the EU pending Digital Markets Act talks.
Reports describe the first confirmed battlefield deaths caused by fully autonomous, AI-controlled drones, after Ukraine is said to have used them to kill Russian soldiers in 2024.
Meta secretly licensed facial-recognition software from Rank One, a Pentagon contractor whose CEO once ran the FBI's biometric-database division, embedded it dormant in the Meta AI app on more than 50 million phones, then deleted it the day after WIRED broke the story.
A KPMG AI study turned into an accidental demonstration of hallucination after GPTZero found only 5 of its 45 citations matched their stated sources, raising questions about how the Big Four firm assembled the report.
PwC reports UK AI-related hiring jumped 61% even as wider vacancies slowed, with employers increasingly seeking workers who can use AI rather than build it.
With coding agents shifting the hard part of engineering from writing code to deciding whether to trust it, a Faros AI study of 22,000 developers found code churn up 861%, the per-developer defect rate up from 9% to 54%, review duration up 441% and zero-review merges up 31%, while GitClear data shows 4x raw output for only about a 12% gain in delivered value.
Leaked audited accounts reported by Ars Technica show OpenAI revenue rising from $3.7B in 2024 to $13.07B in 2025 (monthly revenue nearing $2B by year end), but dwarfed by R&D climbing from $7.81B to $19.18B and an operating loss widening from $8.78B to $20.92B, with a headline net loss near $39B including a roughly $30B one-time accounting charge. The numbers land as OpenAI preps an IPO and weighs price changes to hold enterprise customers.
Days after its record IPO, SpaceX's market value swung up to about $2.6T, at one point overtaking Amazon, underlining how heavily public markets are pricing in its AI and compute potential.
Anthropic shelved a billing change, planned for the following Monday, that would have split Claude Agent SDK usage from standard Claude subscriptions and charged it at higher API rates, sharply raising costs for power users. After pushback, and following recent rival price moves, it said it would revise the plan and let existing users keep current limits for now.
Microsoft patched a critical M365 Copilot vulnerability, dubbed SearchLeak by Varonis, that used a parameter-to-prompt-injection in a crafted URL to make Copilot read and exfiltrate users' emails and data, including 2FA codes, slipping past existing guardrails. Researchers framed it as another sign that the industry's LLM-security approach keeps failing because models cannot reliably separate trusted instructions from injected content.
Stanford's DeLM lets agents coordinate directly through a shared knowledge base and task queue rather than a central controller, working in parallel and passing back only verified findings and failures. It reports cutting multi-agent inference costs by about 50% and beating SWE-bench baselines by 10.5%, targeting the communication bottlenecks of orchestrator-led designs.
A Sina Weibo team's 3-billion-parameter VibeThinker-3B reportedly matches or beats far larger models such as Gemini 3 Pro and DeepSeek V3.2 on reasoning, scoring 94.3 on AIME 2026, prompting both excitement about compact specialised models and accusations of "benchmaxxing". The work argues reasoning can be decoupled from factual knowledge, hinting small models could own narrow reasoning tasks while large ones carry knowledge.
A widely shared post by Vicki Boykis (1,200+ points on Hacker News) argued that open-weight models running locally are now genuinely good enough for everyday coding and assistant work, a marker of how far small and mid-size local models have closed the gap on hosted frontier models.
A report circulating on Hacker News says GitHub is leaning on AWS capacity as demand from Copilot and agentic workflows outstrips Microsoft's own compute, a sign of how acute AI capacity pressure has become even inside the largest clouds.
Alibaba's Qwen team released Qwen-Robot Suite, a set of foundation models for "physical world intelligence" aimed at perception and control for robots, extending the open-model race from language and coding into embodied AI.
A martinfowler.com article drawing on a Bayer engagement set out patterns for making LLM and agentic systems reliable in production, treating non-determinism as a first-class engineering problem with evaluation, guardrails and feedback loops rather than one-off prompt tuning.
TechCrunch reported ChatGPT's share of the consumer AI assistant market fell under 50% for the first time, as Gemini and other rivals chip away at OpenAI's once-dominant lead.
The DOJ moved to block a Clean Air Act lawsuit over xAI's unpermitted gas turbines powering its Memphis compute, arguing the build is a matter of national, economic and energy security and that the military needs Grok, framing AI data-centre power as a defence priority.
The Pentagon publicly touted using AI to draft reports that Congress legally requires, an early example of generative AI moving into core government work and raising questions about accuracy and accountability for statutory documents.
Over the weekend the picture filled out around the export-control ban on Fable 5 and Mythos 5: Amazon's Andy Jassy was confirmed as the "trusted partner" whose calls to officials escalated into a Commerce Department directive, even though the flagged "jailbreak" amounted to Fable finding and helping patch known software bugs rather than producing working exploit code, which is what the guardrails actually target. The fight is as political as technical (David Sacks cast Anthropic as a hypocritical safety company, Pete Hegseth noted the Pentagon ejected Anthropic three months ago), a thinly sourced claim that a China-linked group had accessed Mythos surfaced and was denied, and Anthropic dispatched senior staff (including co-founder Tom Brown and security researcher Nicholas Carlini) to Washington while 40 cybersecurity leaders led by Alex Stamos signed an open letter urging the controls be lifted. This continues the launch-week saga logged 11-14 June.
New York Attorney General Letitia James served OpenAI with a subpoena on behalf of a 42-state coalition (12 Jun), the broadest state investigation yet of an AI company, demanding records on its advertising, engagement-maximising design, consumer and health data, treatment of minors and seniors, and "sycophancy". It lands days after OpenAI's confidential IPO filing (and Anthropic's reported roughly $965B filing) and piles onto a Florida lawsuit that names Sam Altman personally.
OpenRouter launched Fusion, an API that fans a prompt out to several models, has a separate model judge each response, then merges them into one answer. A panel of DeepSeek V4 Pro, Kimi K2.6 and Gemini 3 Flash scored 64.7% on a Perplexity deep-research benchmark against Fable 5's 65.3% for about half the spend, with CEO Alex Atallah pitching it as a bet that "the future of AI is neurodiversity, not single-model takeovers". It also shipped Subagents for handing smaller tasks to cheaper models mid-answer.
Z.ai's new flagship GLM-5.2 reached GLM Coding Plan users, billed for strong coding, usable 1M-token context and continued long-horizon strengths, with API and chatbot access due next week and an MIT-licensed open-source release to follow.
Xiaomi detailed MiMo-V2.5-Pro-UltraSpeed, a 1-trillion-parameter model engineered for 1,000 tokens per second on an 8-GPU commodity node, achieved by codesigning the model with the stack (FP4 quantisation, the DFlash block-parallel speculative decoding method, and startup Tile AI's TileRT runtime). It reads as part of a wider Chinese push to squeeze efficiency out of available hardware under export controls.
Cognition, the maker of Devin, released FrontierCode, 150 graded coding tasks across several languages that judge real code quality (correctness, tests, scope discipline, conventions) rather than just pass or fail. Claude Opus 4.8 manages only 13.4% on the hardest "Diamond" tier (GPT-5.5 6.3%), with Fable reportedly near 30%, and author Jack Clark expects 70%+ on Diamond only around 2027.
Alignment researchers from the UK AI Security Institute and theory startup Timaeus formed Sequent, a nonprofit pursuing a portfolio of theory-led alignment bets (scalable oversight, learning theory, game theory, personas) aimed at higher-confidence safety before recursive self-improvement. It plans 40-80 staff and an initial $100-150M raise, arguing that labs' reactive methods give no principled guarantee of when they will fail and that an independent body "might need to yell".
DeepMind researchers published a paper charting how AI could move from AGI, one broadly human-level system, to ASI, a system or collective that outperforms large expert organisations, through scaling, new model paradigms, recursive self-improvement or massive multi-agent systems.
Microsoft CEO Satya Nadella argued that durable advantage comes from the "token capital" a company builds by feeding its own data and results back into AI over time, and that a frontier without an ecosystem around it is not stable. In a fuller essay covered by VentureBeat he paired this with "human capital" and warned that a handful of frontier models could commoditise expertise and hollow out entire industries, much as outsourcing once did, urging firms to build learning systems that survive a change of underlying model.
OpenAI set up a $150M Partner Network with consulting firms Accenture, McKinsey, BCG, Bain and PwC, aiming to train 300,000 AI consultants by year end to accelerate enterprise adoption.
Google researchers are assembling a data centre from 2,000 retired Pixel phones, due this autumn, to cut the carbon cost of buying new servers. Google Research described the underlying "low-carbon computing platform from retired phones" on its blog.
Disney, the BBC, Adobe, the New York Times and Condé Nast banded together in a new coalition pressing for responsible AI, copyright protection and safeguards for creators and children, with a stated aim of stopping AI from appropriating trusted characters.
A Cornell study reported that roughly 13 words of planted text are enough to steer the outputs of AI deep-research agents, with ChatGPT and Gemini likely similarly vulnerable.
Google is adding a "Skills Marketplace" to Gemini Enterprise with pre-defined, Google-optimised skills plus a Skills Builder and a management UI, pitched at letting teams build dashboards and reporting tools without long engineering queues.
An analysis ("The Oracle and the Firm") contrasts the two labs' approaches to context: OpenAI compacts everything into one long, coherent thread, while Anthropic splits work across sub-agents that each keep their own context window and pass back only the relevant findings, a design that trades some duplicated work and lost detail for parallelism.
Nvidia's Blackwell Ultra NVL72 platform led the new AgentPerf benchmark, delivering about 20x more agent throughput per megawatt than the Hopper generation.
McDonald's is piloting ArchIQ, a Google-powered AI drive-thru system, at five locations, two years after shelving an earlier AI test that produced viral wrong-order clips.
Chinese universities have cut more than 12,000 programmes over five years, shifting away from arts and languages towards technology fields as AI reshapes the graduate job market.
InfoQ covered Anthropic's account of how Claude generates and runs its own "execution harnesses", the scaffolding code that lets a model drive tools, tests and environments, a window into how agentic coding systems bootstrap the structure they need rather than relying on fixed harnesses.
Tenet Security detailed Agentjacking, an attack that plants malicious instructions in fake error events submitted to a project's Sentry DSN, so that AI coding agents reading those errors as trusted guidance execute attacker commands and reach development secrets. Tenet reported an 85% success rate in testing and many exposed DSNs, urging teams to assess agents' combined permissions.
New Relic released an open-source tool for observing AI coding workflows, extending observability practice to track how coding agents behave and perform, part of a wider push to monitor agentic systems rather than treat them as black boxes.
GitHub dropped the personal-access-token requirement for agentic workflows, smoothing how AI agents authenticate to act in repositories and CI, a small but telling step in making agent-driven development the default path rather than a bolt-on.
A VentureBeat essay argued that letting agents "vibe code" data pipelines produces working systems whose intent and design are quickly lost, leaving teams unable to reason about or safely change them later, a concrete restatement of the intent-debt problem for data engineering.
NewCore, founded by Zohar Alon, raised $66M to manage AI agents as first-class identities alongside human employees, with dedicated permissions, controls and a "split-key" authentication architecture, arguing existing identity platforms cannot handle the coming scale of agent access as firms such as Goldman Sachs and McKinsey deploy agents as workers.
Nothing here matches that. Try a broader term, another day, or clear the topic filter.