Google's Gemini 4 Carbon Leads Benchmarks for Finding, Exploiting, and Patching Vulnerabilities

thehackernews.com Oct 31, 2026

Google’s Gemini 4 Carbon Leads Benchmarks for Finding, Exploiting, and Patching Vulnerabilities

Curated by Shubham · TechCapsules · Oct 31, 2026

AI-summarised brief · reviewed before publication

Google unveiled Gemini 4 Carbon, a new frontier model optimized for cybersecurity tasks. The model achieved top scores on three key benchmarks: 88.2 % on SEC‑bench Pro for vulnerability discovery, 54.8 % on ExploitGym for exploit generation, and 71.0 % on CWE‑bench for patch remediation. On the Artificial Analysis Intelligence Index, Carbon led with a score of 59, edging out Anthropic’s Claude Opus 5.5, OpenAI’s GPT‑6 Astra and GPT‑6.1 Sol. Carbon also produces secure code at twice the rate of its predecessor Argon and resists indirect prompt‑injection attacks by deferring to human input when ambiguous. The rollout begins with a limited group of cyber defenders under Google’s Fairwind Program, with broader access planned after U.S. government pre‑release testing. Concurrently, OpenAI released GPT‑6.1 Terra, which meets its own cybersecurity preparedness standards, while Anthropic disclosed an internal model that uncovered and helped fix critical flaws in major software products.

💡 Why It Matters

· Gemini 4 Carbon demonstrates that AI can now autonomously perform the full lifecycle of cyber‑risk management, shifting defensive capabilities from specialist teams to scalable, model‑driven tools.

AI Capsules Google

Read full article on thehackernews.com

Ollie, a privacy-focused AI assistant, is shutting down

Ollie, a privacy‑focused AI assistant startup, announced it will cease operations on November 6, a week after co‑founder and CEO Bill Lennon posted the notice on X. The company, which raised a $7.5 million seed round led by Khosla Ventures, offered a subscription‑based, text‑based assistant that integrated with users’ email and calendar to manage appointments, meal planning, bill payments, and later household budgeting. Ollie distinguished itself by achieving SOC 2 compliance, a rare privacy audit for consumer apps. Despite these differentiators, the firm struggled to compete with a crowded market that includes rivals such as Muse and Instinct, and was unable to secure additional capital. Users will receive instructions for data export and refunds.

💡 Ollie’s shutdown underscores how even strong privacy credentials cannot offset the capital intensity and rapid consolidation shaping the AI assistant sector.

techcrunch.com · Oct 31

Nvidia weighs playing the cosigner for Nscale’s AI data center in West…

Nvidia is negotiating to provide a residual‑value guarantee for billions of dollars in debt financing that AI cloud provider Nscale needs to build Anthropic’s first data center in West Virginia. The guarantee would obligate Nvidia to step in if Nscale defaults on construction payments, take temporary control of the facility and assist in finding a buyer or new tenant. Anthropic has signed a $44.6 billion contract for the initial 460‑megawatt building, which will use Nvidia’s upcoming Vera Rubin chips, and a second roughly $45 billion deal is under discussion. Nscale, aiming for a $35 billion IPO valuation, has struggled to deploy its planned 500,000 AI chips and has faced local construction setbacks. Nvidia’s support expands beyond typical lease and chip‑purchase assistance, deepening its financial exposure to a neocloud that has already received $1 billion of convertible financing from the chipmaker.

💡 Nvidia’s guarantee ties its balance sheet directly to a high‑risk data‑center build, signaling that chip makers are becoming de‑facto financiers to keep AI infrastructure pipelines moving.

therobotreport.com · Oct 31

Percepto launches Axion AI platform to unify inspection data analysis

Percepto, the Israeli‑U.S. drone‑inspection firm, unveiled Axion, a hardware‑agnostic AI platform that shifts the company to a software‑first model. Axion ingests inspection data from any third‑party drone, fixed or mobile sensor and legacy archives, then applies more than 160 AI models to deliver asset‑level condition insights and plain‑language query results. The SaaS service is aimed at critical‑infrastructure operators in oil‑and‑gas, power transmission and similar sectors, allowing them to prioritize maintenance without purchasing Percepto hardware. New CEO Julien Bellanger, who replaced Dor Abuhasira, said the move leverages the firm’s autonomous‑drone expertise while opening its intelligence to the broader market. Percepto will continue to sell its drones and will manufacture them in the United States to support domestic supply chain resilience and growth.

💡 Decoupling AI analytics from specific drone hardware lets operators monetize existing inspection assets, fast‑tracking AI‑driven reliability across legacy infrastructure.

thehackernews.com · Oct 31

Chinese Hackers Use AI-Powered ARTEX to Target Over 200 Organizations Worldwide

Chinese actors identified as Vapor Panda have deployed an ARTEX to compromise more than 200 organizations worldwide since October 17, 2026. The tool integrates DeepSeek and a model, orchestrating planner and worker agents that autonomously conduct reconnaissance, vulnerability discovery and lateral movement across web applications, cloud services, containers, Active Directory and VPN critical infrastructure. CrowdStrike confirmed post‑exploitation at nine victims, where the group installed the BLOODALCHEMY malware, a ShadowPad variant, and carried out credential‑spraying and network scanning. Human operators intervene only to direct agents to create web shells, harvest domain‑admin passwords and evade endpoint defenses. The global campaign targets foreign government ministries, semiconductor firms, defense contractors, think tanks and telecom providers across Latin America, Europe, Asia, the Middle East and the United States.

💡 By automating multi‑vector attacks with AI, a single adversary can breach dozens of high‑value targets simultaneously, reshaping the scale and speed of state‑backed cyber espionage.

windowscentral.com · Oct 31

OpenAI sociopath Sam Altman says it’s “terrible” that your PC, Xbox, etc.…

OpenAI CEO Sam Altman publicly lamented that the memory capacity of consumer devices such as personal computers, Xbox consoles and similar hardware is “terrible” because it could be repurposed to run his company’s artificial‑intelligence models. In a recent interview, Altman argued that the idle RAM in these devices represents a missed opportunity for distributed AI processing, suggesting that a network of everyday machines could collectively provide substantial compute power for large language models. He did not disclose any concrete plans to tap this resource, but the comment sparked discussion among developers and gamers about privacy, security and the feasibility of crowdsourced AI workloads. Altman’s remarks highlight OpenAI’s ongoing search for scalable, cost‑effective compute solutions beyond traditional data‑center clusters.

💡 Altman’s call to harness consumer‑grade RAM hints at a shift toward decentralized AI infrastructure, potentially reshaping how massive models are trained and deployed.

thehackernews.com · Oct 31

Webinar: How to Build AI Threat Readiness Into Your Security Operations

Security teams have long chased faster threat detection, but generative AI is now compressing the window defenders have to act. Advanced models enable attackers to locate vulnerabilities, craft exploit code and traverse networks quicker than defenses anticipate. A webinar titled “How to Build AI Threat Readiness Across Your Security Operations,” hosted week by Wiz, will teach how to fuse disparate data—vulnerability scans, cloud alerts, identity signals, telemetry and threat detections—into a unified context that highlights exploitable paths and prioritizes remediation. The session promises frameworks for improving visibility across cloud, code, SaaS assets, assigning ownership, and shortening the cycle from detection to fix without automating decisions. Attendees will leave with actionable steps to assess and harden their operations against AI attacks.

💡 Faster AI‑driven exploitation means only teams that can instantly translate raw alerts into prioritized, actionable risk will keep pace with automated attackers.