AI Daily Dev — September 2026

Aurora Ransomware Ran Cursor Agent and Claude Sonnet to Hit 20+ Orgs

Sep 1 CloudSEK + Gambit Security disclosures: Russian-speaking Aur0ra affiliates used SpaceX's Cursor Agent, running Claude Sonnet in thinking mode, to plan and drive post-compromise activity.
20+ victims across 9 countries between April and July 2026; 28 recovered chat sessions cover 10 orgs from Apr 8–May 21.
Operators fed the agent SOCKS tunnels and stolen VPN creds, then let it handle lateral movement over SMB/LDAP/WinRM/RDP, log-wiping, Defender kills, and Zig-compiled ESXi locker deploy.
Guardrails bypassed by telling Cursor the engagement was a red-team 'simulation'; prompts explicitly excluded CIS ranges and CIS-country domains.
Lands one day after yesterday's OpenAI-cuts-off-Cursor story and gives xAI a headache heading into Grok 4.7 — Cursor Agent is the first mainstream coding tool to show up in a ransomware IR report.

industry thehackernews.com

Aug 28 CNBC/Bloomberg reporting: DeepSeek is finalizing a ~50 B RMB / $7.4B raise at a $74B pre-money valuation — roughly a 10× step-up on its last mark.
Anchor investors: High-Flyer (Liang Wenfeng's quant fund) plus returning names Monolith Management, Shixiang, Tencent, JD.com, NetEase, and battery giant CATL.
Proceeds earmarked for model research and additional compute ahead of a Shanghai Star Market IPO planned for 2027.
Puts DeepSeek in the top three Chinese AI labs by valuation next to Moonshot and Zhipu, and above Alibaba's Qwen unit on paper — funding round #2 after its debut $7B round in June.

industry cnbc.com

Sept 1 release: same $10/$50 base pricing as Fable 5, but cache reads slashed 75% to $0.25/M tokens — Anthropic says ~25% cheaper on typical workloads, up to 45% on heavily agentic ones.
Terminal-Bench 4.0 = 55.8% (vs 42.0% for Fable 5, 52.3% for Opus 5); Terminal-Bench-Science 0.1 = 52.6%, more than double Fable 5's 24.7%.
1M-token context, 128k output; Mythos 5.1 is the same model under a looser safeguard profile for vetted cyber/biosec orgs, ships alongside a new Enterprise Frontier Safeguards architecture and ~60% fewer cybersecurity false positives in Claude Code.
HN launch thread hits 884 points / 836 comments; Simon Willison's animated-pelican-on-a-bike test rates the high-effort tier the strongest Claude he's used.

models anthropic.com

Sept 2 release: 1M-token context, 64K output, tuned for long-horizon coding and agentic workflows; $0.75/M input and $3.75/M output through Dec 31, both double on Jan 1.
Terminal-Bench 2.1 climbs to 90.8% (81.6% for 3.7 Flash) and HLE-Verified hits 54.9%; beats Opus 5 on three of Google's published benchmarks, though the SWE-Bench Pro gain is barely a point.
3.8 Flash Cyber ships alongside behind Google Fairwind limited access for governments and trusted partners; Chrome Security team says it produces 2.6× more correct patches than comparable commercial models.
HN launch thread splits: Google's DevRel frames higher token usage as 'verifies its work more often', developers counter that 3.8 Flash burned 120M output tokens on a benchmark suite where the median was 71M — 70% more spend at $3.75/M.

models blog.google

Sept 3 announcement: ~$11.9B to investors plus a $1B equity retention pool for HF staff; close targeted for early 2027 pending regulatory approvals.
HF hosts 3M models, 1M apps, and 500K datasets used by 18M developers and 200K companies — the largest open-source AI hub going to a single hardware vendor.
Jensen pledges HF stays 'compute-agnostic': no Nvidia hardware requirement, multi-cloud and multi-accelerator support continue, founding team stays on.
HN thread: 740+ points, 300+ comments — splits between 'Microsoft-buys-GitHub replay' fears and reproducibility upside; awkward timing after last week's OpenAI postmortem on the HF production-environment breach.

industry blogs.nvidia.com

Sept 3 launch: Brockman closes the press briefing with 'Welcome to the AGI era'; 1,050,000-token context, 128K output, knowledge cutoff Apr 30 2026.
Astra saturates FrontierMath Tier 4 v2 at 97.6%, ARC-AGI-3 at 99.9%, and ExploitBench at 100%; 72.6% on OSWorld 2.0 at ~47% less time per task than GPT-5.6 Sol.
Independent Artificial Analysis Intelligence Index lands at 61 — identical to Sol, behind Anthropic's Fable 5.1 — and on Humanity's Last Exam with tools Astra scores 57.2% vs Fable 5.1's 65.0%.
API pricing $10/$50 per M tokens (2.5× Sol); Altman apologizes Sept 4 for a 'messy' rollout after Pro/Plus users were locked out at launch.

models openai.com

Anthropic ran a swarm of parallel Claude agents that produced the first end-to-end machine-checked proof in 11 days of wall-clock time.
Output: 13M lines of Lean 4, 30,300 theorems (29,500 used in the final proof), roughly 6B output tokens.
First formalization attempt failed; adding Columbia's open-source Prove2Me tool mid-run unblocked completion. Lean verified with only its three standard axioms.
Imperial's Kevin Buzzard calls it an 'extraordinary autoformalization achievement'; mathematicians had expected the Wiles proof to take years to formalize.

research anthropic.com

Sept 8: precomputed AlphaGenome predictions for all ~9B possible single-nucleotide variants in the human genome, packaged as a ~1 PB dataset — roughly 30× the AlphaFold DB.
Introduces the AlphaGenome Variant Impact (AVI) score summarising each variant across regulatory, splicing, and expression tracks.
Live via a public website, the AlphaGenome API, and a first-party skill in Google Antigravity; free for non-commercial research.
The largest single-shot precomputed biology asset any frontier lab has shipped this year.

research deepmind.google

OpenAI told reporters an unreleased model, running ~10,000 coordinated agents over 88 hours, produced a Lean-formalized proof of finite-time blowup for the forced 3D Navier–Stokes equations.
Run stats: ~2.7M messages, ~130B output tokens; the model described as 'significantly more capable than GPT-6 Astra.' No proof or model artefact published.
Same day, NYU's Tristan Buckmaster and Anthropic's Levent Alpöge posted a Claude-assisted, machine-checked proof of finite-time blowup for the 3D Euler equations — with the actual Lean files.
HN and math-Twitter latched onto the contrast: one claim in a press call, one verifiable artefact on arXiv.

research unite.ai

Sept 8 joint advisory (AA26-251A) names DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI for systematically distilling Claude, GPT, Gemini, and Grok since late 2024.
Alleges billions of tokens across millions of queries were pulled via fraudulent accounts, bulk premium subs, and proxy 'transfer stations' — likely with Chinese government knowledge.
Specifics cited: DeepSeek R1/V3 trained on Claude, GPT, and Gemini output; Moonshot's Kimi K3 trained on Claude Fable, Kimi K2 on GPT-4o.
Unusual mitigation ask: US labs 'quietly degrade' suspected accounts rather than block them, to preserve the counter-distillation signal.

industry cisa.gov