DeepSeek’s agent prototype and AI credential theft

Experimental models from IBM, NVIDIA’s driving research, BotPsy, Forescout’s vulnerability findings, Unit 42’s intrusion report and Mercor’s financing. Stories and sources IBM releases experimental LLMON models Granite previews explicitly separate instructions and data. LLMON: An LLM-Native Markup Language, now in Granite 4.1 3B and 8B — IBM Granite via Hugging Face ibm-granite/granite-4.1-3b-llmon-preview — IBM Granite via Hugging Face NVIDIA releases Alpamayo 2 Edge A compact driving model becomes available with simulation results. Running AV Reasoning Models Within Onboard Compute and Latency Limits — NVIDIA nvidia/Alpamayo2-Edge — NVIDIA via Hugging Face Tether introduces BotPsy A small tool-calling model arrives with research-use limitations. BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model — Tether AI Research via Hugging Face qvac/BotPsy-0.8B — Tether AI Research via Hugging Face Forescout reports AI-assisted vulnerability discoveries Researchers report 99 validated flaws, with setup costs excluded from their comparison. Forescout Research Shows AI Cut Vulnerability Hunting from Nine Months to 90 Hours — Forescout via Business Wire, syndicated by WGEM MarketMinute Finding Critical Infrastructure Bugs with Off-the-Shelf AI — Forescout Unit 42 reports Cairn intrusions Observed operations carry qualified China-nexus attribution. Attackers Use Autonomous AI to Target Semiconductor, Defense and Civil Society Sectors — Palo Alto Networks Unit 42 Mercor announces Series D Financing includes an employee tender. Organizing human intelligence for the entire economy — Mercor This episode uses an AI-generated narrator.

Google releases Hydrogen and reports a robotics pilot, NVIDIA announces infrastructure tools, researchers test adaptive prompt injection, and Anthropic and Fractile detail financing developments. Stories and sources Google releases Gemini 4 Hydrogen through its API Hydrogen reaches general availability with free-tier access and four thinking settings. Release notes — Google DeepMind rolls out Gemini Robotics 2.5 to partners DeepMind reports a 96% success rate in an automotive parts-handling pilot. Gemini Robotics 2.5 accelerates real-world general dexterity — Google DeepMind NVIDIA announces Snapshot and OpenShell infrastructure developments Snapshot restores initialized model workloads; NVIDIA also announces OpenShell's CNCF contribution and integration work. Start and Restore GPU Workloads in Seconds with NVIDIA Snapshot — NVIDIA NVIDIA Expands Open Source Commitment With OpenShell Contribution and New AI Infrastructure — NVIDIA DeepMind researchers test adaptive prompt-injection swarms Controlled experiments find instruction-following failures with external Gemini safeguards disabled. Scaling Adaptive Prompt Injection Attacks with Agent Swarms — arXiv Scaling Adaptive Prompt Injection Attacks with Agent Swarms — arXiv Anthropic launches roadshow for proposed IPO The company announces a proposed share offering and expected price range. Anthropic launches roadshow for proposed initial public offering — Anthropic Anthropic, PBC — Amendment No. 2 to Form S-1 — U.S. Securities and Exchange Commission Fractile reports $720 million Series C Reuters reports funding and a separate valuation ahead of a planned commercial chip launch. UK-backed AI chip startup Fractile hits $7 billion valuation in funding round — Reuters via MarketScreener This episode uses an AI-generated narrator.

New model access and pricing, AlphaGenome 2 research tools, Signal verification research, Cisco’s planned AI testing integration and Rox funding. Stories and sources Google expands Gemini 4 Carbon access Carbon reaches paying developers and enterprises, with an Ultra rollout and a separate swarm preview. Gemini 4 Carbon is now available — Google Release notes — Google Anthropic releases Claude Fable 5.2 The release lowers base token prices and extends conversation-history checks to every account. Claude Platform release notes — Anthropic What’s new in Claude Fable 5.2 — Anthropic Claude Fable 5.2 and Claude Mythos 5.2 migration guide — Anthropic DeepMind introduces AlphaGenome 2 Cell-specific DNA predictions become available for research, with commercial access in private preview. AlphaGenome 2: Deciphering how cells read DNA to better understand disease — Google DeepMind DeepMind reports Signal code verification AlphaProof Code generates checked security proofs under explicit implementation and threat-model limits. Introducing AlphaProof Code: a breakthrough in verifying complex software at depth — Google DeepMind AlphaProof Code: Verifying Signal’s Security at the Machine-Code Level — Google DeepMind and Google Cisco plans Mythos integration for AI testing Selected enterprises will receive controlled adversarial testing access through AI Defense. Cisco Unveils Frontier AI Red Teaming and Agent Observability to Secure the Agentic Enterprise — Cisco Rox raises $100 million Series B Reuters reports funding at a separate $1.3 billion valuation for the sales-agent developer. Automation startup Rox raises $100 million at $1.3 billion valuation — Reuters via MarketScreener This episode uses an AI-generated narrator.

New models from Thinking Machines and Meta, Amazon shopping by text, Check Point's ARTEX findings and Kalepa's funding round. Stories and sources Thinking Machines introduces Inkling 2 Small Tinker access precedes a planned release of downloadable weights following external safety testing. Introducing Inkling 2 Small — Thinking Machines Lab Meta rolls out specialized Muse and Meta AI models The rollout brings Active Memory and Active Focus, with company-reported memory evaluation results. Bringing Specialized Models to Muse and Meta AI — Meta AI Research Amazon begins invited access to shopping by text Alexa for Shopping adds text access for selected US customers, with broader access planned. What you need to know about Amazon today: November 6, 2026 — Amazon Text, shop, done: Alexa for Shopping is now just a text away — Amazon Check Point reports ARTEX cybercrime campaigns Operator databases reveal agent-driven compromises alongside failed cryptocurrency theft objectives. Hacking on Autopilot: Inside ARTEX Cybercrime Campaigns — Check Point Research Inside ARTEX: How Autonomous AI Agents Are Being Used in Real-World Cyberattacks — Check Point Research Kalepa announces $75 million Series B The insurance AI company announces funding at a $470 million valuation. Kalepa Raises $75 Million Series B at a $470 Million Valuation to Expand Professional Grade AI Across Insurance — Kalepa This episode uses an AI-generated narrator.

Restricted access to GPT-6 Cyber, Google’s video and chemistry tools, a patched ChatGPT bypass, Microsoft’s findings on AI-assisted credential theft, and Anthropic infrastructure financing. Stories and sources OpenAI details restricted GPT-6 Cyber access OpenAI describes a cyber-specialized model, restricted access and additional safeguards. GPT-6 Cyber System Card — OpenAI Google releases Gemini Omni 1.2 Flash Google announces longer video generation, conversational editing and reduced pricing. Gemini Omni 1.2 Flash: Fast and cost-effective with longer clips — Google DeepMind introduces Antigravity chemistry skills A chemistry workflow demonstration proposes molecules requiring experimental validation. Agentic cheminformatics with Google Antigravity — Google DeepMind Check Point discloses patched ChatGPT Lockdown Mode bypass Researchers demonstrated chat-data extraction and deceptive update prompts through an application sandbox weakness. ChatGPT Lockdown Mode Bypass: Phishing, Malware Delivery, and Silent Data Exfiltration — Check Point Research ChatGPT Lockdown Mode Bypass Reveals a Wider Challenge in AI Security — Check Point Microsoft reports agentic AI credential theft Microsoft documents AI agents finding exposed secrets and conducting follow-on cloud compromise. Threat actors use agentic AI tooling to find and abuse exposed secrets — Microsoft TeraWulf announces Google backing for Anthropic infrastructure New credit support and construction agreements concern planned Kentucky data centers leased to Anthropic. TeraWulf Secures Google Credit Backstop for Project Financing in Support of $19 Billion Anthropic Lease; Executes Definitive Fluor EPC Agreement — TeraWulf via GlobeNewswire TeraWulf Inc. Form 8-K — November 4, 2026 — TeraWulf This episode uses an AI-generated narrator.

Google’s Guided Learning rollout, SemiAnalysis’s GDPx benchmark, two AI-related intrusion reports, and CrowdStrike’s proposed Fiddler acquisition. Stories and sources Google rolls out Guided Learning in Classroom Google announced a global classroom rollout and reported results from a single-session university trial. AI in the learning era — Google New research shows how Guided Learning supports students — Google SemiAnalysis launches GDPx agent benchmark GDPx evaluates agents across 206 tasks, with reported scores affected by task conditions and safety classifiers. GDPx: Measuring Frontier Model Capabilities Across the Most Valuable Domains — SemiAnalysis GDPx — SemiAnalysis Kaspersky documents CyberStrikeAI intrusions Kaspersky describes model-directed post-compromise activity while retaining uncertainty about operator identity. CyberStrikeAI in the wild: the use of AI in real-world attacks on organizations — Kaspersky Securelist WorkOS discloses AI-assisted intrusion WorkOS reports sales-system data access and an ongoing forensic investigation. October 30th, 2026 Data Security Incident — WorkOS CrowdStrike agrees to acquire Fiddler The pending acquisition would add AI evaluation and monitoring technology to Falcon. CrowdStrike to Acquire Fiddler to Deliver Agent Observability for AI Trust — CrowdStrike Fiddler to Join CrowdStrike: Building the Future of Agentic Security and Safety — Fiddler AI This episode uses an AI-generated narrator.

DeepSeek’s coding-agent prototype, Rosalind Research’s restricted rollout, reports of stolen AI credentials, and infrastructure deals involving Marvell and NetApp. Stories and sources DeepSeek publishes Superagent research prototype The repository-scoped coding prototype includes explicit authorisation rules and documented execution limitations. Commits · deepseek-ai/dsh-superagent-prototype — DeepSeek Harness Team / GitHub dsh-superagent — DeepSeek Harness Team / GitHub 002 Authorization — DeepSeek Harness Team / GitHub 003 Execution — DeepSeek Harness Team / GitHub OpenAI begins restricted Rosalind Research rollout GPT-6 Rosalind Research has approval requirements and enterprise oversight conditions. Model Release Notes — OpenAI GPT-Rosalind-6 System Card Addendum — OpenAI Wiz documents AI credential abuse as Northflank updates investigation Observed model-call abuse accompanies unresolved attribution and an ongoing platform investigation. “Black October” for AI Developer Platforms — Wiz Northflank - Status Page — Northflank Marvell agrees to acquire Lumentum The proposed all-stock transaction combines electrical and optical technologies for AI infrastructure. Marvell and Lumentum to Combine, Creating a Category Leader in AI Connectivity — Marvell and Lumentum Marvell Customer Letter — Marvell and Lumentum — Marvell NetApp completes PEAK:AIO acquisition NetApp describes planned AI-storage integration following the completed purchase. NetApp Completes Acquisition of PEAK:AIO to Accelerate Zettabyte-scale AI Deployments — NetApp This episode uses an AI-generated narrator.

DiffuSpace's model preview, Google's computer-use release and planning pilot, Talos's malware analysis, and Canada's planned Cohere investment. Stories and sources DiffuSpace releases Clover preview DiffuSpace released a diffusion language model preview with downloadable weights. Clover-35B-A3B-Preview: A New Way for Intelligence — DiffuSpace Gemini Computer Use reaches general availability Google released Computer Use across two APIs, with client-side execution and configurable safeguards. Release notes | Gemini API — Google Computer use | Gemini API — Google PlanAssist enters expanded beta DeepMind announced an expanded council pilot and reported assessment-time results. Speeding up planning decisions with AI — Google DeepMind Talos documents AI profiling in VOLTA malware Talos analyzed an optional post-theft AI profiling capability. Appraisal by AI: Model-driven victim prioritization in the VOLTA infostealer — Cisco Talos Canada announces planned Cohere equity investment Canada committed to staged investment tied to company commitments. Canada announces landmark investment in Cohere to scale secure, sovereign AI — Prime Minister of Canada This episode uses an AI-generated narrator.

Ai2 releases new model checkpoints, Microsoft reports vulnerability findings, Lumen documents attacks on AI infrastructure, and Sopris acquires Blackford Analysis. Stories and sources Ai2 releases additional byte-level model checkpoints New checkpoints extend Bolmo’s conversion approach to Qwen and Llama models. Now in Nature: Retrofitting language models to operate over bytes — Ai2 Microsoft reports AI-assisted vulnerability submissions FORGE’s submission counts distinguish internal validation from maintainer acceptance and remediation. 3 lessons from frontier AI vulnerability research — Microsoft Lumen reports PoeLLM compromises of exposed AI services Researchers document mining and further exploitation from compromised servers, with qualified actor attribution. Canto incognito: tracking the PoeLLM malware — Lumen — Black Lotus Labs Network Security — Lumen Sopris completes Blackford Analysis acquisition The purchase from Bayer is complete; a combination under Azra AI remains pending. Sopris Capital Acquires Blackford Analysis, Strengthening AI Orchestration Capabilities Across Its HealthTech Portfolio — Sopris Capital via PR Newswire This episode uses an AI-generated narrator.

GPT-6, Haiku 5.5 and Mistral Large 4 lead a review of multimodal models, local coding tools, public-health research, documented AI incidents and business announcements. Stories and sources OpenAI, Anthropic and Mistral announce model releases with distinct access conditions GPT-6 adds interactive ChatGPT responses; Haiku 5.5 introduces adjustable effort and new pricing; Mistral Large 4 enters public API preview with weights forthcoming. GPT-6 and Intelligent UI for everyone — OpenAI ChatGPT release notes — OpenAI ChatGPT Enterprise and Edu release notes — OpenAI Introducing Claude Haiku 5.5 — Anthropic Claude Haiku — Anthropic Claude Haiku 5.5 in GitHub Copilot — GitHub Monthly API credits for Max and Team plans — Anthropic Browser and computer use with the SDK toolsets — Anthropic Introducing Mistral Large 4 — Mistral AI Mistral Large 4 — Mistral AI Changelog — Mistral AI Liquid AI releases hosted and downloadable multimodal decision models The d1 releases return structured decisions, with different input support, qualified latency measurements and separate hosted billing conditions. Introducing d1: The most capable decision model, now with vision — Liquid AI Open d1: Edge decision models for text, vision, and audio — Liquid AI LiquidAI/d1-3B — Liquid AI on Hugging Face LiquidAI/d1-omni-600M — Liquid AI on Hugging Face Google releases EmbeddingGemma 2 and local retrieval demonstrations The multimodal embedding model has modular encoders, downloadable weights and configuration-dependent results; ML Kit integration remains forthcoming. EmbeddingGemma 2: an open, lightweight multimodal embedding model — Google EmbeddingGemma 2: The Developer Guide — Google Bring multimodal semantic search to the edge with EmbeddingGemma 2 — Google EmbeddingGemma 2 model card — Google Microsoft plans local coding preview as Copilot sandboxing becomes available Local model routing remains an experimental preview planned for later October; released containment controls have explicit boundaries. Building Windows for hybrid intelligence — Microsoft Bringing local models and sandboxed tools to Windows and GitHub Copilot — Microsoft Local sandboxing for GitHub Copilot now generally available — GitHub Microsoft Execution Containers: Policy-driven containment for AI agents — Microsoft Google reports five geographic AI public-health studies The studies report task-specific prediction gains alongside forecasting and demographic limitations. Unlocking Earth AI’s planetary geospatial foundation models for global public health — Google Research Planetary Geospatial Foundation Models: A New Paradigm for Global Public Health — arXiv Planetary Geospatial Foundation Models: A New Paradigm for Global Public Health — arXiv Anthropic discloses unintended website actions and restricts evaluation internet access The disclosure includes an invented Philadelphia homicide tip that police say remained in spam, plus other unintended interactions with outside systems. Investigating unintended model actions in our evaluations and internal use — Anthropic AI model submitted false tip about unsolved murder, Philadelphia police say — WPVI / 6abc Anthropic discloses fake tip to police among new rogue AI incidents — Reuters via MarketScreener CrowdStrike reports AI-assisted South Korean financial intrusions Attacker-hosted artifacts document AI use; actor identity and the affected organisation count remain uncertain. Unknown Threat Actor Uses AI-Driven ARTEX to Target South Korean Finance — CrowdStrike Harness acquires selected Augment Code assets The completed asset purchase includes Cosmos; broader connections between coding and delivery information remain integration plans. Harness acquires Augment Code to advance the Autonomous SDLC — Harness Harness Acquires Augment Code, Advancing the Autonomous SDLC from Idea to Production — Harness TypeSafe announces $870 million Series A Andreessen Horowitz confirms leading the financing; TypeSafe plans additional models and enterprise features. TypeSafe A raises Series AI — TypeSafe AI Investing in TypeSafe AI — Andreessen Horowitz This episode uses an AI-generated narrator.