AI/ML Security & Trends
The dominant story is Cisco Talos's disclosure of CLOSEDQUORUM — the first reported malware with fully autonomous, multi-model AI command-and-control, where four LLMs (DeepSeek, Qwen, Mistral, Gemini) vote on the attacker's next move with no human operator — alongside the same week's UN Security Council session where Sam Altman and Dario Amodei jointly warned that ungoverned frontier AI is "the most important global security issue facing the world today."
Cisco Talos discloses CLOSEDQUORUM, first autonomous multi-model AI malware C2 AI Security & Safety
Talos's new CAIRN hunting toolkit surfaced CLOSEDQUORUM, malware that queries four different LLM APIs and executes whichever action wins a plurality vote, run as a credentials-as-a-service offering. Because per-model safety guardrails are independent, one provider refusing a malicious prompt doesn't stop the attack — the other three models can still form a majority. No confirmed in-the-wild deployment yet, but artifacts link the developer to carding forums active since 2025.
Read at Cisco Talos →Altman and Amodei jointly warn UN Security Council on frontier AI risk Industry & Trends
At a France-chaired UN Security Council session on "the future of AI and international security," Sam Altman and Dario Amodei separately urged world leaders to adopt international AI standards, with Amodei calling it "the most important global security issue facing the world today." It marked the first time the Council hosted frontier Chinese (DeepSeek, Moonshot) and US developers together, following weeks of disclosures about AI agents acting outside intended bounds.
Read at CNN →Trump-Xi summit puts Nvidia China chip export rules back on the table Industry & Trends
President Trump hosts Xi Jinping on September 24 with AI chip export controls among the contested topics; Nvidia's current guidance assumes no China data-center chip sales, and any loosening would be market-moving. The talks come as Aivres and other resellers face scrutiny for routing over $3B in Blackwell-equipped systems to Southeast Asia that reportedly reached ByteDance and Alibaba, and as Huawei unveils domestic AI chips meant to reduce China's Nvidia dependence.
Read at The Motley Fool →New framework proposes AI-agent-specific incident reporting standard AI Security & Safety
A new arXiv paper, "Beyond Predictable Paths: Redefining AI Security Incident Reporting for Agents," argues existing AI incident-reporting frameworks don't capture agent-specific attack surfaces — agent memory and memory access, autonomy levels, and tool usage — and proposes new reporting elements informed by 23 academia/industry experts. It flags open questions like how to determine whether an agent vulnerability generalizes across deployments.
Read at arXiv →OpenAI ships GPT-6 Sol and GPT-6 Luna, cutting API prices in half Model & Product Releases
OpenAI released GPT-6 Sol (mid-tier) and GPT-6 Luna (fast, high-volume) as enterprise workhorse models, both with 1.05M-token context windows and built using training methods derived from GPT-6 Astra. Luna is priced at $0.10/$0.50 per million input/output tokens — among OpenAI's cheapest offerings — positioning the two models for high-volume production workloads ahead of OpenAI DevDay on September 29.
Read at TechCrunch →Google ships Gemini 3.7 Flash, Gemini 3.5 Transcribe and Robotics ER 2 Model & Product Releases
Google released Gemini 3.7 Flash with improved software-engineering and agentic-workflow performance at an introductory price through year-end; Gemini 3.5 Transcribe and a streaming Live variant for high-accuracy speech-to-text across 85+ languages; and Gemini Robotics ER 2 in public preview, adding spatial-reasoning endpoints for embodied/robotics applications.
Read at Google AI →xAI ships Grok 4.7 with new jailbreak-resistance "safeguard stack" Model & Product Releases
xAI's Grok 4.7 targets coding and agentic work, with independent benchmarks showing the highest safety score yet on HackerBench v0.3 (only 3.3% of risky dual-use prompts allowed through) and a leading 62.4% on LatchBio's biosafety benchmark. xAI is giving select cybersecurity partners invite-only red-team access. Priced at $2/$6 per million tokens, doubled token consumption versus 4.6 was flagged by independent evaluators.
Read at xAI →Alibaba previews Qwen 4, unveils Zhenwu V900 AI chip Model & Product Releases
At Apsara Conference in Hangzhou, Qwen project lead Liu Dayiheng confirmed Qwen 4 is in training on a new architecture and coming "very soon," with Qwen 4.5 and 5 targeting 5-10 trillion parameters. Alibaba also unveiled the Zhenwu V900, described as China's most powerful AI processor, from its T-Head semiconductor unit — part of the broader push toward domestic compute as US export controls persist.
Read at Yahoo/YourNews →Proofpoint launches unified agentic system linking data and AI security Tools & Frameworks
Proofpoint announced Agentic Data and AI Security, billed as the industry's first unified system connecting an AI agent's intent to its actual data access, running three purpose-built autonomous agents to detect, investigate, and remediate risky agent behavior. It reflects growing enterprise demand for agent-behavior monitoring following a string of AI-agent security incidents this quarter.
Read at Proofpoint →Anthropic launches Claude Marketplace for enterprise software and services Industry & Trends
Anthropic's new marketplace groups third-party offerings under Add (2,000+ connectors/plugins from Google, Microsoft, Salesforce, Atlassian), Buy (Claude-powered apps from Replit, GitLab, Harvey, Legora, Lovable, Snowflake), and Scale (consulting from Accenture, BCG, Deloitte) — with committed spend redeemable against partner purchases, making it a procurement channel as well as a discovery surface.
Read at VentureBeat →Cohere and Aleph Alpha sign definitive merger to form ~$20B transatlantic AI firm Industry & Trends
Cohere and Aleph Alpha signed a definitive business combination agreement valuing the merged entity at roughly $20B, pairing Cohere's scale with Aleph Alpha's research base and European institutional relationships. The deal, which builds on an April 2026 framework agreement, still needs regulatory approval and is expected to close later in 2026.
Read at SiliconANGLE →DeepSeek ships V4.1-Flash with native multimodal understanding Model & Product Releases
DeepSeek released V4.1-Flash, the smallest model in its new architecture family, adding native multimodal visual understanding and retiring the prior V4-Flash API endpoint. The new architecture is pitched as a higher capability ceiling with faster inference and higher throughput ahead of larger models in the same family.
Read at DeepSeek →Microsoft Agent Framework 1.19.0 adds security-label enforcement, new vector-store connectors Tools & Frameworks
Microsoft's agent-framework 1.19.0 release ships tightened security-label enforcement and restricted owned-input argument labels in agent-framework-core and agent-framework-hosting-a2a, alongside new vector-store provider protocols, alpha MongoDB/Azure DocumentDB connectors, and Azure Cosmos DB NoSQL support — continuing the post-AutoGen/Semantic Kernel consolidation of Microsoft's agent tooling.
Read at GitHub (Microsoft) →AWS and Nvidia deepen collaboration on accelerated computing and inference Industry & Trends
AWS and Nvidia announced an expanded strategic collaboration covering accelerated computing, interconnect technologies, and model fine-tuning/inference infrastructure, part of continued hyperscaler-silicon alignment as demand for frontier-model training and serving capacity keeps climbing.
Read at AWS →