AI News and Industry Updates

The Same Class of Sandbox Flaw: Cursor and OpenAI Fixed It in About a Week, Anthropic Took Roughly 50 Days and About 30 Releases

Accomplish, a security startup still in stealth (founded by Or Hiltch, Amit Avner and Guy Zipori), publicly disclosed leaky-sandbox problems in Claude Code, OpenAI Codex and Cursor after privately notifying the vendors over the summer. The newsworthy part is not the vulnerabilities but the difference in patch cadence: one flaw reported to Cursor in July and two reported to OpenAI were all fixed in roughly a week, while a similar issue reported to Anthropic about two months earlier went unpatched for some 50 days, during which the product shipped around 30 updates. Accomplish CEO Amit Avner argues attackers could have been exploiting the gaps during that window — his assessment, with no evidence of actual exploitation given in the reporting. The founders describe this as a "long tail" of lesser-known, almost normalized vulnerabilities that startups and researchers keep surfacing. In the same window, Manifold Security, Pillar Security and Cymulate each disclosed trust-boundary problems of a similar nature in coding agents.

AI Content Editor

Cognition Post-Trained on Kimi K3 and Got a Coding Model Within a Point of Fable 5.1 at 64% Less Cost

Cognition released its coding model SWE-2 on September 10, post-trained from Kimi K3, Moonshot's 2.8-trillion-parameter open-weight model. The company's figures: 50.0% on FrontierCode 1.1 Main, within one point of Fable 5.1 while costing 64% less; 73.0% on DeepSWE 1.1 and 92.8% on Terminal-Bench 2.1, gains of 5.8, 4.5 and 4.5 percentage points respectively over Kimi K3's original scores. The methodological point worth remembering is effort-level training: this is Cognition's first model with reasoning effort levels, all trained in a single RL run with a linear cost penalty per level, each penalty tuned to the local slope of the base model's Pareto frontier. The efficiency gains are concrete — SWE-2 medium scores higher than the previous SWE-1.7 on FrontierCode 1.1 Main while taking 58% fewer turns and costing 81% less on average, and the median step at which it makes its first real edit drops from 48 to 18. The weak spot is in Cognition's own materials too: 27.3% on Terminal-Bench 4, against 57.9% for GPT-6 Astra. It is available in Devin Desktop and CLI, free for Pro/Max/Teams subscribers for one month, through October 10.

AI Content Editor

Hundreds of Agents Compromised 11 Organizations in 26 Seconds — but Only 12 of 440 Breached Instances Reached Domain Admin

GreyNoise published a report titled "Agents Gone Wild": starting August 31, an attacker first used AI in a self-hosted lab to develop and test exploits for two PaperCut NG/MF vulnerabilities (CVE-2026-81578 and CVE-2026-82078), then — once remote code execution and credential harvesting worked — deployed hundreds of AI agents at scale, using OpenAI's Codex as the harness, a DeepSeek model as the driving model, publicly available offensive security tooling, and target lists built with the internet scanning service Netlas.io. Speed is the report's most striking part: under four hours from an empty workspace to first RCE against a real victim, another two hours to first domain admin, and once the campaign launched, at least 11 organizations compromised in 26 seconds — with one US high school going from initial access to domain admin in seven minutes. At least 440 instances across 395 identified organizations in 48 countries were affected, education hardest hit at 204; yet while credentials were harvested on 280 instances and OS or domain secrets on 147, domain admin was reached on only 12. PaperCut issued emergency patches on August 28.

AI Content Editor1 views

Amazon Can Buy Up to $60 Billion of Qualcomm Data-Center Silicon, and Part of the Payment Structure Is a $4 Billion Warrant

On September 8, at a Goldman Sachs investor summit in San Francisco, Qualcomm and Amazon disclosed a long-term partnership under which Amazon can purchase up to $60 billion of Qualcomm AI data-center chips, systems, technology and manufacturing services over the next decade, focused on inference. The structure is unusual: Qualcomm issued Amazon warrants to acquire 25 million shares at $161.26 each, $4 billion in total, vesting in phases as Amazon engages commercially, places binding orders and actually takes delivery of equipment — with 3.75 million shares vesting immediately on the first pledges and the warrant expiring in September 2036. Bernstein analyst Stacy Rasgon estimates the immediately vested portion alone implies an initial sales commitment of roughly $9 billion. The two will also work across multiple generations of customized silicon and co-develop high-speed optical connectivity carrying up to 1.6 terabits per second inside AWS data-center networks, using Qualcomm's SerDes and optical DSP technology; reciprocally, Qualcomm will increase its use of AWS and Amazon Bedrock to shorten chip design cycles. Qualcomm has already received purchase orders and begun production, expects to recognize revenue from December, and targets $15 billion of data-center chip revenue by 2029.

AI Content Editor

Gemini Finally Has a Windows Desktop App: Alt + Space Summons It, and That Shortcut Is Its Only System-Level Hook

Google released the Gemini desktop app for Windows on September 10, about five months after the Mac version, as a free download from gemini.google/desktop. It is not a browser tab but a standalone program: Alt + Space by default pulls a floating window over whatever you are running, and the shortcut can be swapped for another preset or a custom one; the side panel holds chat history, Daily Brief, image and video creation, document libraries, Notebooks and recent creations. Google AI subscribers can hand tasks to Gemini Spark, the personal agent, with image generation via Nano Banana and video via Gemini Omni. A free account signs in to a base model with daily usage limits, while higher usage and stronger models such as Gemini 3.1 Pro require Google AI Plus, Pro or Ultra — AI Pro listing at $19.99 a month. Google says the app is lightweight and won't slow the machine down, with more native desktop capabilities rolling out over time.

AI Content Editor
GitHub AI Weekly · Issue 05

GitHub AI Weekly · Issue 05 (Sep 12, 2026) · AI Beyond

This issue covers GitHub Trending for ISO week 37 of 2026, ranking ten eligible AI projects by stars gained this week across AI coding, agents, multimodal work and AI infrastructure. Every entry explains the week's change, who it suits and where it falls short, with three deeper analyses. Data is current through Sep 12.

AI Content Editor1 views

Meta Bought an Eight-Month-Old Swedish AI Company, and Its Promise to "Keep Serving Existing Customers" Carries No Date

Axios reported exclusively on September 9 that Meta is acquiring Swedish AI startup Stilla.ai to accelerate Meta Business Agent — the agent that completes transactions for merchants inside WhatsApp, Messenger and Instagram — with terms undisclosed. Stilla was founded in Stockholm in 2024 by former Shopify executives Siavash Ghorbani and Kaj Drobin, disclosed a $5 million pre-seed when it left stealth in January, and builds AI agents that work across an organization's tools and workflows with an emphasis on shared context, permissions and enterprise-grade infrastructure; the team and technology join Meta once the deal closes. Stilla says its platform will continue operating as a service for existing customers, but that pledge carries no end-of-support date, no connector maintenance schedule and no commitment from Meta itself, and the deal was still pending close when reported. Meta Business Agent launched globally at the Conversations conference in London on June 3 and is now used by more than one million businesses.

AI Content Editor1 views

OpenAI Turned the Harness Behind Codex Into an API: the Agents API Enters Public Beta, Owning Sessions, Context Compaction and Recovery

OpenAI opened the Agents API in public beta on September 10, exposing the orchestration layer that previously served only Codex and aiming it at tasks that run for minutes, hours or days. What it hosts is sessions, orchestration, context compaction and recovery from interruptions; you supply the tools and choose the execution environment. Four concepts define the model: Agent (model, instructions, tools, MCP servers), Environment (an optional sandbox for files, skills and commands), Session (a durable agent instance) and Events/items (inputs and outputs). Built-ins include automatic context compaction (performed independently for the root agent and each subagent), tool search that loads tool definitions on demand, programmatic tool calling that runs calls in parallel or chained and filters results in code, and subagents with isolated context. Sandboxes can be OpenAI-hosted, your own, or from Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel. There is no fee for the API itself — you pay normal model rates plus tool and container usage. Two hard limits: it is United States only, and it is not eligible for zero data retention, even with a self-hosted sandbox.

AI Content Editor

OpenAI Published Its Navier–Stokes Proof and the Lean Formalization — and Says It Won't Claim the Millennium Prize, Because This Isn't the Classic Version

OpenAI announced a result on the Navier–Stokes existence and smoothness problem on September 8, then on September 10 published the full technical write-up, the paper and a machine-checkable Lean formalization, saying it will not claim the Clay Mathematics Institute's million-dollar Millennium Prize. The scale figures are specific: nearly 10,000 agents working in parallel reached the result on September 5, about 88 hours after the first agents launched, sending 2.7 million messages and producing roughly 130 billion output tokens, after which GPT-6 Astra spent another 17 hours on Lean formalization and verification. The boundary has to be stated plainly: what was proved is that the 3D Navier–Stokes equations develop a singularity under a smooth forcing term applied throughout, which OpenAI frames as establishing statements C and D of the official Millennium Prize formulation rather than the classic problem itself. Separately, NYU mathematician Tristan Buckmaster has raised a priority dispute that OpenAI denies; the two accounts do not currently agree.

AI Content Editor

Oracle's Contracted Backlog Hit $664 Billion. Free Cash Flow in the Same Quarter Was Negative $5 Billion

Oracle reported first-quarter fiscal 2027 results on September 10: total revenue of $19.35 billion, up 30% year over year; cloud infrastructure revenue more than doubling, up 121% to $7.4 billion; total cloud revenue up 62% to $11.6 billion; and net income available to common shareholders up 60% to $4.68 billion. The headline number is remaining performance obligations, which rose $209 billion year over year to $664 billion, above the $630.6 billion consensus, with more than $30 billion of additional AI cloud contracts booked in the quarter. The delivery figures are equally concrete: 850 megawatts of AI capacity and over 300,000 GPUs delivered in the quarter, close to triple the prior quarter, at 97.9% GPU utilization. The other side of the ledger: $28 billion of capex in the quarter produced negative $5 billion of free cash flow, full-year capex is guided at $90–95 billion, and Oracle completed a $20 billion at-the-market equity sale during the period. Gross margin declined as expected on data-center ramp, with management anticipating a further step down.

AI Content Editor

Every Possible Single-Letter Change in the Human Genome, Precomputed: DeepMind Ships the 1-Petabyte AlphaGenome Atlas

On September 8, Google DeepMind released AlphaGenome Atlas: using the 2025 AlphaGenome model, it precomputed the molecular effects of all 9 billion single-nucleotide variants in the human genome, producing roughly 1 petabyte of data — more than 30× the AlphaFold Database (the genome has about 3 billion base pairs, and each position allows three possible substitutions) — plus over 100 million short insertions and deletions observed in human genomes. Alongside it comes the AVI score, which fuses AlphaGenome and AlphaMissense outputs into one disruption ranking spanning both coding and non-coding regions, with feature attributions and 2,500+ motifs explaining why a variant scores high. Early validation involved the Broad Institute, the University of Exeter and the Stowers Institute: a validated DNM1 splice variant was identified, and 22% more non-coding associations surfaced across 54,000+ UK Biobank genomes. The non-commercial web portal is free and the API is on GitHub, while commercial use via Google Cloud is "coming soon" with no pricing disclosed; DeepMind states the output is not a substitute for professional medical advice and is not approved for clinical use.

AI Content Editor

Legal AI Is Now Priced at Nearly 40× ARR: Harvey Raises Another $550M — and the Money Is Going Into Building Its Own Models

On September 9, legal AI company Harvey announced a $550 million round; its own blog puts the valuation at $15.5 billion (Bloomberg reported $15.6 billion, a difference likely reflecting pre- versus post-money), co-led by Diffusion and Lightspeed Venture Partners, with Sapphire Ventures and Whale Rock joining and existing backers including Sequoia, Kleiner Perkins, a16z, Coatue, GIC and Goldman Sachs Alternatives participating. ARR has passed $400 million and the customer base has grown from 1,300 in March to more than 3,000 organizations, including Latham & Watkins, Microsoft's in-house legal team, 80% of the Am Law 100 and five Fortune 10 companies; total funding now exceeds $1.5 billion. The valuation path runs $3 billion in February 2025, $5 billion mid-year, $8 billion in December, $11 billion in March 2026 and $15.5 billion now — roughly 39× revenue. The company says the money mainly goes toward building proprietary models, and it also acquired AI agent security startup Guardrails AI this week.

AI Content Editor

Showing 12 of 12 stories