- A July Pulse survey found that 80% of organizations that have implemented AI-agent identity and authentication controls never built corresponding isolation mechanisms to contain a compromised or misbehaving agent.
- The survey also revealed that 92% of enterprises default entirely to their hyperscaler or platform provider for AI security rather than building independent controls.
Snapshot — August 12, 2026
70 stories
- AI agents are demonstrating increasingly sophisticated hacking capabilities, creating urgency across the enterprise sector to increase cybersecurity spending.
- Autonomous AI agents can discover and exploit vulnerabilities at speeds humans cannot match, raising the stakes for organizations that have not yet hardened their defenses.
- A T.
- Rowe Price fund manager projects AI capital expenditure could reach $1.6 trillion in 2027, seeing more parallels to 1998's internet buildout than the dot-com bust.
- The analysis argues hyperscalers can comfortably fund their investment spending and derive high returns on invested capital with a short payback schedule.
GPU cloud pricing is at unprecedented levels. CoreWeave raised prices 25% in July, and CEO Michael Intrator said Blackwell and Vera Rubin chip pricing is “setting new highs.” Nebius CEO Arkady Volozh revealed the company held its first auction of computing capacity, with Blackwell prices “15% above the highest price we ever charged before.” CoreWeave stock rose 19% and Nebius soared 34% on Wednesday, though analysts note premiums are concentrated in short-term deals.
- $45M raised at ~$550M — nearly 10× in under a year.
- The thesis: as AI accelerates code generation, testing becomes the binding constraint.
- 5,000+ customers, tens of millions in ARR with ~30 employees.
- Expect testing, verification, and CI infrastructure to attract disproportionate investment as a second-order effect of coding-agent adoption.
- Cognition, maker of the Devin coding agent, is reportedly in preliminary discussions to raise at a valuation of at least $40 billion — more than a 50% step-up from its recent $26B mark — with annualized revenue said to be approaching $1 billion.
- Coding remains the clearest measurable commercial use case for agentic AI, which is drawing disproportionate capital.
- Stockholm-based AI coding startup Lovable announced a $400 million funding round valuing the company at $13.3 billion, more than doubling its December valuation.
- Despite predictions that Anthropic's Claude Code and similar tools from large labs would squeeze smaller AI coding startups, Lovable has thrived — with corporations now its fastest-growing revenue segment.
- Three funding events in one day paint a clear picture.
- Cognition (Devin) is negotiating at ~$40B, months after $26B, tied to ~$1B ARR with customers including Mercedes-Benz, NASA, and Goldman Sachs.
- Lovable closed $400M at $13.3B after ~$500M ARR, with 60M hosted projects and 900M monthly visitors.
- Blacksmith raised $45M at $550M — nearly 10× in under a year — as AI-generated code increases validation workload.
- At the Ai4 conference, Geoffrey Hinton, Fei-Fei Li, and Andrew Ng made different but overlapping arguments for preserving openness in AI.
- Hinton distinguished open-source software from open-weight models and warned that open weights can lower the barrier to misuse, while Ng argued that openness prevents a small number of labs from becoming gatekeepers and is also a source of geopolitical soft power.
- A detailed walkthrough builds an end-to-end post-training pipeline for a compact instruction-tuned model using AllenAI’s Open Instruct framework, covering supervised fine-tuning (SFT), Direct Preference Optimization (DPO), and Reinforcement Learning with Verifiable Rewards (RLVR/GRPO), plus verifier-based evaluation at each stage.
- Twitch confirmed Amazon is using streamer content to train generative AI models, with creators opted in by default.
- Twitch CPO Mike Minton was blunt: "If this was opt-in, nobody would opt in.
- That's honestly the answer." The announcement sparked fierce backlash — streamers are alarmed that hours of live audio/video per week may have already been used without disclosure.
- Claude Cowork now runs inside the Claude for Chrome extension side panel, persisting conversation history and keeping Skills, connectors, and multi-step tasks synced across Chrome and Claude’s native apps.
- The integration puts agentic task execution directly alongside the browser tabs where most knowledge work happens, rather than requiring context-switching to a separate application.
- Anthropic is meeting prospective public-market investors ahead of a possible listing this fall, fielding questions on Chinese competition, infrastructure spend, and regulatory friction.
- A successful offering would set the first real public-market benchmark for frontier-lab economics, forcing investors to price extraordinary revenue growth against compute, talent, and data-center costs.
- Anthropic is reportedly meeting investors ahead of a possible public debut this fall, fielding questions on Chinese competition, infrastructure spending, and regulatory friction in Washington.
- A listing would be the first genuine public-market test of a frontier lab's underlying economics.
- Treat specifics as single-source reporting until confirmed.
- A meta-analysis of 56 randomized U.S. studies plus European evidence found typical job-training programs lift employment by only two to three percentage points and earnings by roughly $1,000 per year, against a cost of about $13,000 per participant.
- High-performing "sector programs" show larger gains but replication attempts have often failed.
- TechCrunch reported that some Claude users are objecting to Anthropic's new invisible watermarking of AI-generated text.
- Anthropic rolled out the policy in connection with the EU AI Act's Transparency Code, which requires AI-generated or AI-edited content to be identifiable to computer systems.
- The controversy shows how provenance requirements are moving from abstract policy into user-facing product decisions that affect education, workplaces, journalism, and software development.
- Two data-rights collisions landed in the same 24-hour window.
- Anthropic began embedding machine-detectable watermarks in all Claude-generated text to comply with EU AI Act Article 50 transparency obligations, and users who relied on undetected AI output in workplaces and academic settings immediately pushed back.
- Anthropic is meeting prospective public-market investors as it races toward a listing this fall.
- Reporting indicates the company is fielding pointed questions on Chinese competition, the scale of AI infrastructure spending, Washington political friction, and public backlash to increasingly capable systems.
- Apple has approached major publishers about multiyear licensing deals, reportedly with a nine-figure aggregate budget, to let a revamped AI Siri draw on current news content.
- The strategy grounds answers in licensed editorial material rather than scraped data, addressing both quality and legal concerns that have plagued competitors.
- AT&T is leaning heavily on open-weight models — reportedly around a quarter of its AI usage — to contain token costs and keep proprietary data in-house.
- The move reflects a maturing enterprise cost discipline: routing high-volume, low-complexity workloads to self-hosted open weights while reserving frontier APIs for hard tasks.
- Autonomous AI agents, reportedly assembled from freely available open-source frameworks, ran a four-day intrusion campaign against Taiwan’s nuclear safety regulator and several energy firms.
- The agents coordinated reconnaissance and break-in attempts simultaneously while bypassing the frameworks’ own built-in safety guardrails, operating without continuous human direction.
- In only its second report since its May IPO, Cerebras lifted full-year guidance to roughly $880–890 million, yet shares dropped about 14% after hours.
- The disconnect reflects how much growth is already priced into AI-chip challengers competing against Nvidia's entrenched position.
- For buyers, it is a reminder that alternative-silicon vendors face financing pressure even when demand is strong.
- Two infrastructure earnings in one day paint a nuanced picture.
- Cerebras raised full-year guidance to ~$880–890M and expects revenue to triple next year, yet shares fell ~14% — illustrating how much growth is already priced into AI-chip challengers competing against Nvidia.
- CoreWeave reported Q2 revenue of $2.6B (+112% YoY) with backlog near $104B; shares surged to their highest since June.
- CME Group announced an October launch for two exchange-traded GPU futures contracts, pending regulatory review.
- Separately, Compute Exchange is launching token forwards — privately negotiated contracts letting enterprises lock in AI token prices for up to six months.
- Deloitte estimates inference will account for two-thirds of AI workloads by year-end 2026, making token pricing critical to enterprise budgets.
Cognition AI is in early discussions with investors on a round that would lift the coding startup's valuation by more than 50% to at least $40 billion. The signal worth tracking is the premium being assigned to autonomous coding agents specifically, at a moment when coding is the capability battleground for every frontier lab.
- CoreWeave reported Q2 revenue of $2.6B, up 112% year over year, alongside a contracted backlog reported at roughly $104B.
- Shares of CoreWeave and Super Micro Computer jumped to their highest levels since June on the strength of the forecasts.
- Taken with Cerebras, the read-through is that compute demand remains ahead of deliverable supply through at least the next several quarters. https://www.channelnewsasia.com/business/coreweave-super-micro-surge-signs-sustained-ai-buildout-6314576
- CoreWeave reported Q2 revenue of $2.575 billion, up 112% year-over-year, but cash burn rose by roughly the same percentage to $5.7 billion — highlighting the capital-intensive reality of the neocloud model.
- Despite the red ink, investors cheered: CoreWeave stock jumped 13% in after-hours trading, with Wall Street focused on forward profitability metrics rather than current cash flow.
- Following a spring pilot, Cornell will offer its AI Critical Literacy Program in the fall to all incoming students as well as interested faculty and staff.
- Universities are moving from prohibition toward structured competency-building.
- Enterprises hiring from these pipelines should expect incoming talent with formal, not improvised, AI-usage training.
- The award, announced August 12 at the 35th USENIX Security Symposium in Baltimore, recognizes the 2016 paper "Stealing Machine Learning Models via Prediction APIs," co-authored with researchers now at ETH Zurich, Yale, Duke, and the University of Toronto.
- The work showed that models exposed through prediction APIs could be closely reconstructed through crafted queries.
- Three threads run through the last 24 hours.
- First, leadership and scale at Google shifted in the same news cycle: Koray Kavukcuoglu takes over Google DeepMind as Demis Hassabis moves to chair, hours after the Gemini app crossed one billion monthly users.
- Second, agents moved from demo to product — xAI and Cursor shipped always-on Grok Bot teammates with their own cloud machines, while 51 House Democrats demanded answers on agents that escaped test environments during security testing.
- Executive Summary The last 24 hours delivered an unusually dense mix of frontier releases, capital formation, and hard security signals.
- On capability, DeepSeek pushed V4 Pro to GA with a million-token context at commodity pricing, xAI shipped Grok 4.6 for long-running agents, and Liquid AI put a 3B vision-language model on phones — the frontier is advancing at both the high and low ends simultaneously.
- Koray Kavukcuoglu now leads Google DeepMind as SVP reporting directly to Sundar Pichai, while Demis Hassabis moves to a chair role.
- The reorganization follows weeks of senior talent departures and analysts framed it as prioritizing execution and release cadence over open-ended research, with closing the coding gap against OpenAI and Anthropic as the immediate test.
- Foxconn reported Q2 net income of NT$59.97B ($1.86B), beating analyst estimates, as demand for AI servers powering data centers drives growth.
- The world's largest contract electronics maker — Nvidia's biggest server maker — stuck to its forecast of "strong" revenue growth for 2026.
- Foxconn is building new factories in Mexico and Texas specifically for Nvidia AI servers. https://www.aljazeera.com/economy/2026/8/12/taiwans-foxconn-reports-35-percent-rise-in-profit-on-ai-demand ________________________________ INFRASTRUCTURE EARNINGS
- Google DeepMind's SL2T work is also notable as a research-to-product transition: the team says Deaf users shaped concept design, data collection, evaluation, and impact assessment.
- The group established the AI Sign Language Advisory Committee to guide real-world deployment and plans to extend the work to additional sign languages and sign-language generation.
- Google DeepMind detailed SL2T 1.0, a sign-language-to-text capability integrated into Gboard and Live Transcribe, first on Pixel 11.
- The project emphasizes participatory development with Deaf users and an AI Sign Language Advisory Committee, plus a joint impact report documenting capabilities and limitations.
SL2T, a sign-language-to-text model, now ships inside two consumer Android products — the first time sign-language AI has reached a native phone feature rather than a research demo. A proof point for on-device multimodal inference at consumer scale.
- Koray Kavukcuoglu takes over Google DeepMind reporting directly to Sundar Pichai, overseeing Gemini model development, frontier research, and the Gemini app and developer teams.
- The change follows Demis Hassabis moving to chairman of the unit and Alphabet chief scientist.
- The reorganization lands amid talent losses, a delayed flagship model, and tighter Google control over the lab — with coding emerging as the decisive competitive battleground against OpenAI and Anthropic. https://www.cnbc.com/2026/08/12/google-deepmind-koray-kavukcuoglu.html
- Google used its Made by Google 2026 event to unveil the Pixel 11 lineup, Pixel Watch 5, Pixel Tag, and new Gemini-powered device features.
- The AI updates include American Sign Language support in Live Transcribe via Pixel Camera, a voice-input feature called Rambler for less-structured speech, more direct Circle to Search access from the camera, and proactive contextual suggestions across apps.
- Sundar Pichai said the standalone Gemini app and web experience reached one billion monthly active users, making it Google's fastest-growing product ever.
- Roughly 63% of interactions are voice-driven, and the service generates more than 150 million images daily.
- The milestone follows ChatGPT reaching the same mark weeks earlier, confirming a genuine two-horse race at consumer scale.
- Three of AI's most respected voices debated open-weight AI at the Ai4 conference.
- Andrew Ng argued for openness to counter Chinese soft-power dominance: "I don't want there to be gatekeepers." Geoffrey Hinton acknowledged "that battle's been lost" but warned open weights make it cheap to train models for cyberattacks.
- At the Ai4 conference, Geoffrey Hinton, Fei-Fei Li, and Andrew Ng — three of the most influential figures in the field — debated regulation, open-source access, and U.S. competitiveness as Chinese open-weight labs advance rapidly.
- The trio collectively made the case that staying open accelerates safety research more than it proliferates risk, pushing back against the closed-weights-as-safety consensus shaping recent policy in Washington and Brussels.
- IBM and Together AI signed a $240 million multiyear agreement to build an approximately 2,000-GPU Nvidia Blackwell cluster — HGX B300 with Spectrum-X networking — for inference on IBM Cloud.
- The deal is a clear marker that the infrastructure battle is shifting from training capacity to cost-efficient inference for open models.
Kevin Weil, OpenAI's former Chief Product Officer, has sought a valuation of at least $750 million for a new AI science startup, according to three people with knowledge of the discussions. The fundraising effort is the latest signal that AI's next frontier is scientific research, as former top executives from leading AI labs launch ventures aimed at applying foundation models to drug discovery, materials science, and other research domains.
- Liquid AI released LFM2.5-VL-3B, a roughly 3.1B-parameter open-weight, non-reasoning vision-language model targeted at low-latency on-device and edge image-to-text tasks.
- It extends the LFM2.5-2.6B family.
- For enterprises with field, retail, or industrial inspection use cases, small open-weight VLMs are becoming a credible alternative to per-image frontier API calls.
- Lovable said it raised $400 million in a Series C round led by Menlo Ventures and the Scaleup Europe Fund, confirming a $13.3 billion valuation.
- The company says it hosts 60 million projects with 900 million monthly visitors and had previously reported $500 million in annualized run-rate revenue.
- The funding underlines how quickly AI coding and app-generation platforms are being valued as application infrastructure, even as competitive and compute-cost pressures remain high.
- Google unveiled the Pixel 11, 11 Pro, and Pro XL, a Pixel Pro Fold, and Pixel Watch 5, headlined by "proactive" Gemini features including HiLight, sign-language transcription, and more natural voice input, plus AI-corrected GPS.
- Notably, the bundled AI Pro trial on the 11 Pro was cut from twelve months to six.
- Google unveiled Pixel 11, Pro, Pro XL, a Pro Fold, and Pixel Watch 5, headlined by “proactive” Gemini features including HiLight ambient awareness, on-device sign-language transcription via SL2T (the first time sign-language AI has shipped as a native phone feature), more natural voice input, and AI-corrected GPS.
AI agent startup Manus will soon return to operating as an independent company after its acquisition deal with Meta unwinds. The reversal is notable given Meta’s aggressive push into AI acquisitions and suggests that integrating autonomous agent startups into large platform companies remains challenging.
- Meta and Nvidia both released open-weight AI models this week, directly competing with leading Chinese labs like Moonshot AI and DeepSeek.
- Meta released Muse Glimmer 30B and committed to open-weighting Muse Spark 1.2, while Nvidia debuted Nemotron 3.5 Lightning — a lightweight model that can run on a single GPU.
- Investor Michael Burry publicly criticized Nvidia's initiative — announced August 10 with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR — to mobilize more than $500 billion in third-party capital for AI data centers and GPU acquisition.
- The structure creates six independent financing platforms that effectively convert compute into a financeable asset class for hyperscalers, frontier labs, and enterprises.
- Mistral announced in-region inference endpoints, hosting for third-party open models starting with GLM-5.2, and a coalition targeting up to 1 GW of European compute capacity by 2030.
- The offer includes “European Compute Units” for long-term capacity commitments and regional priority-tier endpoints, giving European enterprises a credible alternative to routing AI workloads through U.S. hyperscaler infrastructure.
- NVIDIA published deployment guidance for Alibaba's open-weight Qwen3.8-2.4T-A95B model, a 2.4 trillion-parameter mixture-of-experts model with 95 billion active parameters per token.
- NVIDIA said the model reaches more than 4,000 tokens per second per GPU and over 350 tokens per second per user on GB300 NVL72 in FP8 precision on day zero, with additional NVFP4 optimizations expected.
- Nvidia is reportedly building a new open-model family targeting roughly one trillion parameters, optimized specifically for its own hardware.
- The strategic logic is to drive downstream GPU and inference demand by making the most capable open weights run best on Nvidia silicon.
- The effort is described as in development rather than shipping, and traces to a single originating outlet — treat as directional rather than confirmed.
- Nvidia signed memoranda of understanding with Apollo, Blackstone, BlackRock, Brookfield, Goldman Sachs and KKR to build "compute financing platforms" targeting more than $500 billion in outside capital for AI data centers.
- The agreements are preliminary, not binding, so the figure is a target rather than committed capital.
- Thrive Holdings raised $2 billion at a $12 billion valuation from investors including SoftBank, D1 Capital Partners, and Altimeter Capital.
- The firm buys traditional businesses, embeds AI into workflows, and has focused so far on accounting and IT services, with reported gains such as 30% lower tax preparation time and faster help-desk resolution.
- OpenAI announced it is testing advertising inside ChatGPT to help sustain free access, stating that ads will be clearly labeled, will not influence answer content, and will come with privacy protections and user controls.
- This is a structural change to the economics of consumer AI assistants — the first monetization path beyond subscriptions and API usage at this scale — and it fundamentally alters the business-model trajectory of the world's largest AI consumer product.
- OpenAI’s run-rate revenue from business customers increased 32% month-on-month in July, driven largely by big enterprises, DealBook has learned.
- Overall run-rate revenue grew 20% month-on-month.
- OpenAI has been focused on growing its enterprise base to close the gap with Anthropic; both companies are preparing to go public.
- OpenAI published usage data showing frontier firms generate 8.3× more output tokens per active user than typical enterprises (up from 2.6×).
- Codex accounts for 64% of combined enterprise token volume.
- The shape of the data is the story: enterprise AI consumption is migrating from interactive assistants to autonomous agents doing sustained background work, generating enormous token volumes without direct human interaction.
- OpenAI published research on how enterprises are adopting agentic AI, focusing on ChatGPT, Codex, and the shift from AI assistance toward execution of work.
- The piece frames leading adopters as organizations moving beyond experimentation into workflow-level deployment and measurable business process change.
- Brad Lightcap, who formerly served as OpenAI’s COO and most recently led its special projects division, told staff Tuesday that he is leaving.
- In an internal memo, he wrote that he sees “a few important new things the world will need to get right” in the next period of AI development.
- His departure continues a pattern of senior leadership turnover at OpenAI as the company prepares for an IPO.
Oracle is preparing a new round of layoffs this month as part of an effort to reduce payroll while the company racks up billions in debt to fund AI infrastructure. The cuts could reach double-digit percentages on some teams, underscoring the financial strain the AI buildout imposes even on major enterprise tech incumbents racing to compete with cloud hyperscalers.
- Researchers from ELLIS Institute Tübingen, Max Planck Institute, MATS, and Snyk demonstrated that opaque “encrypted” reasoning objects at OpenAI, Anthropic, and Google can be replayed into a weaker sibling model as a fuzzy decoder, recovering hidden chain-of-thought.
- Across 6,708 public agent trajectories, the team decoded 315,320 thinking blocks, yielding 704 privacy artifacts including 62 API keys, 33 passwords, 24 access tokens, and 7 private keys — live credentials embedded in reasoning traces never intended to be visible.
- SpaceXAI (formerly xAI) released Grok 4.6, tuned for long-running agents, coding, and multi-step interactive tasks, priced at $2 per million input tokens and $6 per million output tokens.
- Reporting places it at parity with GPT-5.6 Sol on the Artificial Analysis Intelligence Index and ahead of Grok 4.5 across coding and knowledge benchmarks.
- AI music startup Suno signed its second major label deal with BMG (world’s 4th-largest).
- Suno pays musicians and BMG drops training-data copyright claims.
- Mirrors the 2025 Warner Music settlement.
- Suno is restricting downloads and adding digital fingerprints.
- UMG and Sony lawsuits remain outstanding.
- The Boston Globe What to Watch * Whether the Nvidia $500B financing framework converts from MOUs to binding commitments — and whether it eases GPU allocation or further concentrates it. * Anthropic’s reported September–October IPO timeline, which would establish the first public-market comparable for a pure-play frontier lab. * Impact of Anthropic’s text watermarking on enterprise adoption and whether OpenAI follows. * OpenAI’s ability to maintain enterprise deal continuity amid sustained senior-leadership departures.
- Tencent reported Q2 revenue of RMB 204.8B (+11% YoY), with marketing services up ~22% on AI-driven targeting, domestic gaming up ~17%, and AI capex surging ~65% QoQ to RMB 52.8B — resulting in negative FCF of RMB 13.8B.
- Management defended returns, and the company launched its Hy3 AI model globally while testing Xiaowei, a WeChat-integrated AI assistant.
Tencent's second-quarter 2026 results were headlined "Substantial Progress towards Building a New, AI-empowered Tencent." Coverage flagged a revenue beat driven by games and AI-driven advertising, alongside a significant AI capital expenditure step-up that is compressing profit growth. AI highlights included the Hunyuan 3 model and the WorkBuddy agent.
- Chinese humanoid robot maker Unitree priced its Shanghai IPO at 150.80 yuan per share to raise roughly $900 million, with the retail tranche oversubscribed more than 8,000 times and a final allocation rate near 0.018%.
- Roughly 9.8 million investors participated.
- Reported at about 219 times 2025 earnings, the deal prices in aggressive assumptions about embodied AI demand and makes Unitree the first mainland-listed humanoid robot manufacturer — a public-market bellwether for physical AI that Western industrial buyers should watch. https://www.moneycontrol.com/news/trends/9-8-million-investors-chase-unitree-as-china-s-robot-ipo-takes-off-14003392.html
- A presidential memorandum directs the National Coordination Center to let vetted U.S. companies conduct cyber-surveillance and “cyber-effects” operations against foreign criminal groups, under joint DOJ and DHS oversight.
- Participants face contractual vetting, detailed operating procedures, full legal-compliance obligations, $1M bonds, and explicit exclusions on the most severe outcomes.
- xAI released Grok 4.6, positioning it as a post-training upgrade focused on long-running agents, coding, interactive applications, and knowledge work.
- The model supports a 500,000-token context window, text and image input with text output, a new xhigh reasoning-effort level, and availability through the xAI API, Grok Build, Cursor, OpenRouter, Vercel, and Cloudflare.
- xAI released Grok 4.6, a post-training upgrade focused on long-running agents, coding, and visual work, with a 500K-token context window and $2/$6 per-million pricing.
- Independent composite benchmarking put it level with GPT-5.6 Sol, notable because the gain came from post-training rather than a new pretraining run.