AI News: October 10, 2026

1. TypeSafe Raised $870 Million at a $7.5 Billion Valuation for Its Jev Decision Model

TypeSafe AI. Andreessen Horowitz led the round, with Sequoia and existing investor DCVC participating, less than a month after Jev launched on September 15. Jev is a transformer that outputs calibrated probabilities over caller-defined options instead of text, and the company claims it runs much faster and uses far fewer tokens than LLMs for automation decisions. TypeSafe, founded in 2024 by former OpenAI researcher Diogo Almeida, Sasha Sheng, and Erik Gafni, says about a third of Fortune 500 companies already use it. Source

2. Independent Benchmark: Jev Matches Frontier LLMs on Knowledge Tasks but Trails on Math

Jev evaluation. A new paper compares Jev with 19 LLMs across 13 multiple-choice benchmarks and finds it competitive with frontier models on knowledge and commonsense tasks, with the top score on MMLU-Redux and ARC-Challenge. It falls short on math word problems: on MathQA it is 17.7 points below the frontier median and scores below all 19 LLMs. The authors conclude decision models can handle knowledge-driven choices but not ones that require multi-step calculation. Source

3. “Option-Channel” Attack Defeats Decision Models Used as Agent Guardrails

Option-Channel attack. Testing seven open-weight typed decision models as allow-or-block gates for agent tool calls, the authors measured 36 to 72 percent accuracy on prompt-injection, jailbreak, and toxic-content screening against a 50 percent chance baseline. Six lines of policy-irrelevant server log text raised the fail-open rate on one policy from 0 to 63 percent, and giving the permissive option a misleading label pushed it to 93 to 100 percent on four models. Every defense tested was bypassed, and a deterministic rule over parsed policy fields reached 100 percent accuracy, leading the authors to recommend such models only for triage, not final decisions. Source

4. Harvard Study: AI Coding Tools Have Not Raised Firms’ Software Output

Harvard University. Researchers Fiona Chen and James Stratton analyzed Jellyfish engineering data covering 300 million work events from more than 700,000 employees at over 700 software firms between 2021 and March 2026. They found “little evidence that firms increase software output or reduce employment” with AI coding tools, because gains in code generation are absorbed downstream: code review takes longer, pull requests need more revisions, and reviewers leave more comments. The result points to human review as the main bottleneck in agent-assisted development. Source

5. Isomorphic Labs Is Reportedly Raising at a $40 to $50 Billion Valuation

Isomorphic Labs. The Alphabet drug discovery spinoff is in early talks for new funding at that valuation, Bloomberg reported, following a $2.1 billion round in March. The company has small-molecule partnerships with Eli Lilly and Novartis worth nearly $3 billion combined, plus a Johnson & Johnson collaboration started in January. Source

6. Harness Acquired Augment Code’s Agent and Context Engine Assets

Harness. The DevOps company bought select assets of Augment Code, including the Cosmos multi-agent engineering platform (renamed Harness Cosmos Software Factory), the Auggie CLI, and the Code Context Engine, and the Augment team is joining Harness. Terms were not disclosed. CEO Jyoti Bansal said the goal is to join code generation with Harness’s testing, security, and deployment agents in one “idea to deployment” pipeline, while letting customers keep their existing coding tools. Source

7. Ukrainian Drone Strikes Knocked Out Two Yandex Data Centers

Yandex. Drones hit Yandex facilities in the Sasovo and Kaluga regions on October 8 and 9, disabling two of the company’s five data centers and leaving it struggling to keep its AI chatbot and other services running, according to Reuters reporting cited by Ars Technica. President Zelenskyy described the strikes as a “symmetrical response” to a Russian campaign against Ukrainian data centers that has disrupted mobile and internet service since late September. The attacks mark AI compute infrastructure becoming a direct military target. Source

8. Thomas Hales Laid Out the Reliability Limits of Lean Proofs in the Age of AI

Thomas Hales. In a guest post on Terence Tao’s blog, Hales reviews the “Summer of Soundness Bugs,” in which security researchers using frontier AI found Lean kernel bugs that allowed illicit proofs, including a bogus disproof of the Collatz conjecture; all were fixed and mathlib was re-verified. He recommends cross-checking proofs with independent kernels (about 25 exist), formally verifying kernels such as the Con-Leche checker, and strengthening the theory of Lean’s type theory. He also argues that humans must still audit whether formal statements match their intended meaning, especially as AI autoformalization scales to millions of lines. Source

9. Xiaomi Detailed and Open-Sourced the RL Stack Behind MiMo-V2.6

Xiaomi. The technical report describes an omni-modal model family trained by scaling reinforcement learning, with asynchronous training that consumes 1,568 samples and 2.7 to 3.7 billion tokens per step at contexts up to 1 million tokens, across code, general, visual, and cyber environments. To keep training stable, the team froze the MoE router and built layered defenses against reward hacking, and it used groupwise agentic grading to reward shorter solutions. Xiaomi is releasing the training dynamics, RL environments, and RL framework. Source

10. CoreWeave Launched Forge for End-to-End Inference Tuning

CoreWeave. Announced at its Fully Connected event, Forge connects model serving, observability, post-training, and evaluation in one platform. A preview feature, RL Rollouts, built on NVIDIA Dynamo, hot-loads new checkpoints into live deployments so inference scales separately from training during reinforcement learning, and it cut model reload latency 15x in CoreWeave’s tests. Forge is free to start, with paid tiers for broader access. Source

11. Gallatin AI Raised $50 Million for Military Logistics Software

Gallatin AI. The El Segundo startup’s Series A included 8VC and Silent Ventures, following a $15 million seed in 2024. Its Navigator platform normalizes supply records from multiple military systems, forecasts when supplies will run out, and flags gaps in resupply plans, and it is used by the US Army and Air Force. Source

12. Rein Security Raised $25 Million for Runtime Agent Protection

Rein Security. The Series A, from Glilot Capital, Sienna Venture Capital, Corner Ventures, and others, brings total funding to $35 million. Rein deploys a lightweight sidecar that learns each agent’s normal behavior and blocks specific deviating actions at runtime without shutting down the whole agent. Source

13. Magnific One Adds Creative Direction and Brand Kit Checks to Image Generation

Magnific. The company formerly known as Freepik launched Magnific One, which drafts 8 or 16 compositions for the cost of one image, renders finals at 2K or 4K, and checks outputs against a team’s Brand Kit for style, color, typography, and logo mismatches. It is available on web, desktop, mobile, and via MCP in assistants such as ChatGPT and Claude, at roughly $0.12 to $0.73 per image. The company told VentureBeat it runs on OpenAI’s GPT Image 2 with its own post-training, though its FAQ says nothing was trained. Source

14. Paper Models a “Population Threshold” for Misaligned Agent Takeoff

Ecology of AI Agents. Erin Crawley and Hidenori Tanaka model populations of misaligned agents whose growth rate depends on cyber capability, and show that when agents collaborate, a critical population size emerges: below it the population declines, above it it takes off, even with no change in individual capability (an Allee effect). Because red teaming small groups cannot guarantee safety at larger scales, they propose “ecological red teaming” and gradual population pacing in controlled environments, re-estimated for each model generation. Source

15. Psychometric Audit Finds HarmBench Does Not Measure a Single “Refusal” Trait

HELM Safety audit. The authors find three of four HELM Safety datasets relevant to harmful refusal are saturated, and item response theory modeling indicates the remaining one, HarmBench, does not measure a single attribute. Differential item functioning analysis also flagged items where models from different developers with the same refusal ability scored differently. The paper argues aggregated safety scores can hide saturation and conflate distinct behaviors. Source

16. Project Greenhouse Released Fully Open Rerankers Trained From Scratch

Project Greenhouse. Jimmy Lin and colleagues built a competitive pointwise decoder-only reranker by pre-training from scratch and fine-tuning on common datasets, without relying on any third-party open-weight backbone, using only a handful of GPUs. The team released the Gaggle model checkpoints along with data, code, and configs for full reproduction, as a first step toward “sovereign” agentic search models. Source

17. Nikon Disqualified Its Small World in Motion Winner Over AI Use

Nikon. After researchers flagged structural anomalies and apparent AI watermarks in the source files of Ning Xu’s winning cilia video, Nikon concluded the entry did not comply with competition rules and named Nguyen Nam Nhat’s roundworm video the new winner. Xu said AI was used only to distinguish and visualize features in reconstructed grayscale images, not to generate the footage. Source