AI news Nov 3: Agent security and engineering models
CrowdStrike made AI agent security the day's clearest business theme with an acquisition, a testing deal and a kill switch. Elsewhere, new systems from Krea, Ai2, Google and Snowflake focused on image editing, long-context reasoning, engineering and autonomous work.
- CrowdStrike combined agent acquisitions with an AI kill switch.
- Krea 3 targets fast, high-quality image editing.
- Ai2's StateBench exposes long-context state tracking problems.
- Google and OpenAI advanced engineering-focused AI.
- Snowflake added autonomous agents and governance controls.
Agent security becomes a platform priority
CrowdStrike plans to acquire Fiddler AI to connect AI agent protection with checks on whether agents behave safely and as intended. The company also announced its intent to acquire Dreadnode, whose technology tests models, applications and agents for security weaknesses before production.
CrowdStrike launched an AI kill switch in Falcon Guardian, allowing teams to cut off model access through AI gateways when agents become dangerous. A separate announcement says Falcon Guardian protection is coming to Anthropic's Claude Cowork, with the sensor running inside its virtual machine. Falcon AI Gateway is now generally available in Falcon Guardian.
Models test the limits of reasoning and creation
Krea published the technical report for Krea 3, an image-to-image model that supports up to 8 reference images and finishes edits within 8 seconds in production, according to the company. Ai2 introduced StateBench to measure state tracking across long documents, finding that recurrent and hybrid models can generalize farther than transformers on these tasks.
Google introduced Physical-Engineering's Last Exam, with 29 public tasks for testing mechanical, electrical, civil engineering and structural design. OpenAI demonstrated models for engineering design and manufacturing, while Google DeepMind showed Google Antigravity generating drug candidates and possible synthesis routes in about an hour.
Enterprise AI adds autonomy and controls
Snowflake added autonomous Cloud Agents and scheduled automations to CoCo. Its related updates include Profiles and Dashboards, Business Ontology and governed context, a central agent gateway and spending limits. Cortex Training entered public preview for serverless fine-tuning of open-weight models.
SAP and Cohere made North generally available on SAP Business AI Platform for sovereignty, data residency and compliance needs. Apple expanded its redesigned Health app to 10 language options and added a Longevity tab. Datadog Agent Observability joined Claude Marketplace, while E2B launched Workload Identity.
Infrastructure, chips and industry deployments
AMD reported $13.4 billion in third-quarter revenue, with data center revenue up 97% year over year. Supermicro is building inference systems combining NVIDIA GPUs with SambaNova SN50 RDUs. Nordic unveiled nRF71, combining connectivity and on-device AI, while MediaTek, Ericsson and T-Mobile US recorded above-5G speeds in a pre-6G test.
Saudi Arabia's AI plans face data center infrastructure constraints. Applied Computing unveiled Orbital 2.0 for energy operations. Other announcements include PATEO's local AI Box, LG Energy Solution's ESS-led quarter and Windward's acquisition of Azimut.ai.
More launches and security findings
Cogent's VR-1.1 fixed 57.9% of vulnerabilities in its evaluation. JFrog found five Reachy Mini vulnerabilities that could enable unauthenticated control on the same network. Other updates include Hugging Face's CWE-bench v2, Resemble AI's Teams deepfake detection, Vercel Security Review and Tencent Hy4 in Vercel AI Gateway.
Aleph Alpha's Kolibri offers four reasoning settings. Daytona reported growth from a $1M to $100M run rate. Hebbia added Office and Copilot Cowork features, Ultima Genomics won a Singapore sequencing award, and Longsys announced an industrial wide-temperature mSSD.
