inno A+I ve

AI Daily (6 Nov) – OpenAI launches GPT‑6 Cyber; Dropbox buys Tigris for $350M

•

Richard L. Z.

Today’s top line: OpenAI unveiled GPT‑6 Cyber for authorised vulnerability testing while companies double down on AI infrastructure and lower creative costs – Dropbox pays $350M for Tigris and Google cuts Gemini video pricing 30%.

OpenAI releases GPT‑6 Cyber to find vulnerabilities within authorised scope

Summary: OpenAI introduced GPT‑6 Cyber, a cybersecurity model designed to find exploitable vulnerabilities and develop fixes while being trained to remain within an authorised scope. The model is presented as a specialised offering for security testing.

Digest: A frontier model aimed at security testing changes how organisations might automate vulnerability discovery and patch suggestions; watch for how enterprises integrate and govern its use.

Source: OpenAI

Dropbox buys Tigris Data for $350M to build DBX Agent Cloud

Summary: Dropbox agreed to acquire cloud storage company Tigris Data for $350 million, a deal expected to close in Q4 2026. The acquisition will back DBX Agent Cloud, an AI-focused storage platform that initially offers S3-compatible object storage and is already reaching more than 70,000 teams; the updated Dropbox experience supports connectors to ChatGPT, Claude, and Muse.

Digest: The purchase signals broader platform bets on AI-native storage and agent continuity for enterprises, and could lower friction for businesses hosting data used by models and agents.

Source: TechCrunch

Google rolls out Gemini Omni 1.2 Flash with 15s video and 30% lower cost

Summary: Google DeepMind released Gemini Omni 1.2 Flash, increasing native generated video length to 15 seconds (up from 10 seconds) and cutting price by 30% to $0.07 per second for 720p. The update, available via the Gemini API, Google Vids and Gemini for business, adds multi-turn conversational editing and stronger reference adherence for characters, products and brand details.

Digest: Lower per-second costs and longer clips make AI video more practical for ads and product explainers; businesses that produce frequent short-form video should reassess budget and workflow trade-offs.

Source: Google DeepMind

Axiom formally verifies the Stacks Project into 16 million lines of code

Summary: Axiom Math said its agents translated and formally verified the Stacks Project – a 7,654-page mathematics library – into about 16 million lines of Lean code. The $1.6 billion startup, which raised a $200 million Series A in March, used proof assistants and agent workflows to preserve the project’s interlinked definitions and theorems.

Digest: Formal verification at this scale could provide a trusted mathematical foundation for future agentic systems that must reason reliably; expect more verification-first tooling for safety‑critical software and research.

Source: TechCrunch

Agility Robotics lists on Nasdaq and details robot pricing and rental plans

Summary: Agility Robotics debuted on Nasdaq and published unit economics for its Digit humanoid robot. The company outlined a robots‑as‑a‑service model with about $25,000 one‑time delivery/deployment and $8,500 monthly fees, and an ownership option with $200,000 hardware cost plus $36,000 annual software and maintenance fees; Digit 5 is slated to ship in 2027 and can lift up to 50‑pound loads.

Digest: Public filings give rare transparency into humanoid robot margins and customer costs, informing how manufacturers and customers will evaluate return on investment for automation in logistics and factories.

Source: Ars Technica

Hugging Face ships Phonon on‑device TTS – 31m model is Apache 2.0

Summary: Hugging Face released Phonon, a small text‑to‑speech family that runs on devices and browsers. Models come in ~65MB and ~75MB sizes, cover five languages with 100+ voices, and Phonon 31m is Apache 2.0; the project saw more than 200,000 downloads in its first two weeks.

Digest: High-quality offline speech that runs on user devices reduces latency and privacy risk and makes voice features cheaper at scale for product teams building embedded or offline experiences.

Source: Hugging Face

Agent Harness Awareness paper shows verified telemetry cuts agent failures 97%

Summary: An arXiv paper introduces Agent Harness Awareness, where hosts supply small verified state snapshots before consequential actions. In a 216‑decision ablation across GPT‑5.6 Sol, Claude Sonnet 5, and Gemini 3.7 Flash, verified tool‑health state reduced mean observable policy regret by 96.9% (4,252 ms to 131 ms) and raised single‑dispatch completion from 69.4% to 95.8%; a live 27‑decision replication cut mean regret from 8,956 ms to 0 ms and completed live actions in 9/9 cases versus 5/9 for the control.

Digest: Supplying small, verified telemetry snapshots before tool use materially improves agent decision accuracy and reliability – a practical design for safer tool‑using agents in production.

Source: arXiv

Conformal Reservation for Orchestration halves resource waste for agentic pipelines

Summary: Another arXiv paper frames uncertainty‑aware orchestration for agentic AI and proposes Conformal Reservation for Orchestration (CRO), an online algorithm that converts recent execution traces into dispatch‑time compute and bandwidth reservations with finite‑sample, distribution‑free coverage via conformal prediction. Simulations show CRO is the only evaluated method to meet probabilistic coverage targets while roughly halving resource over‑reservation versus a Chebyshev‑bound baseline.

Digest: CRO offers a provable way to reserve resources for unpredictable agentic pipelines, lowering wasted capacity while reducing the risk of mid‑execution throttling – a key operational tool for running agentic systems at scale.

Source: arXiv

Quick hits

•