AI News 11.03: Agent Security and Autonomous AI Systems
On November 3, the security and autonomy of AI agents took center stage. CrowdStrike is expanding its defenses through several announcements, while new benchmarks, data platforms and physical engineering applications point to the industry's next direction.
- CrowdStrike is expanding its AI security portfolio by acquiring Fiddler AI and Dreadnode.
- The Falcon Guardian AI emergency shutoff can block model access for agents that become dangerous.
- StateBench and Google's engineering benchmark measure new capabilities.
- Snowflake offers autonomous data agents and serverless model fine-tuning.
- A Reachy Mini vulnerability could enable remote control without authentication.
CrowdStrike: Acquisitions and Immediate Protection for Agents
CrowdStrike plans to acquire Fiddler AI to bring together AI agent security, monitoring and behavioral evaluation. The company also plans to acquire Dreadnode. Its automated testing technology would assess AI models, applications and agents on the Falcon platform before they are deployed in production.
Falcon Guardian customers will receive an AI emergency shutoff that can quickly cut off model access for agents that become dangerous. The company also introduced a dedicated AI Kill Switch controller, while also integrating its protection into Anthropic's Claude Cowork virtual machine.
New Benchmarks and Autonomous Data Tasks
Ai2's StateBench benchmark measures how language models track information that changes across long documents. The test found that recurrent and hybrid architectures generalize further with direct training. Google, meanwhile, is using the Physical-Engineering's Last Exam benchmark to assess capabilities in mechanical, electrical, civil and structural engineering.
Snowflake's CoCo system can assign complex data tasks to autonomous Cloud Agents, even after the user has left, while retaining existing permissions and policies. The company's Cortex Training service offers serverless fine-tuning for open-weight models.
Models for Engineering Work and Healthcare
OpenAI is testing its models on mechanical and electrical design, simulation and manufacturing tasks. After analyzing thousands of compounds, Google DeepMind's Antigravity system proposed new macrocycles and synthesis routes in about one hour.
Apple's new Health app is expanding with ten language versions, personalized Apple Intelligence insights and at-home mobility assessments. Aleph Alpha's Kolibri model chooses from four levels of reasoning effort, while Resemble AI has launched real-time deepfake detection for Microsoft Teams.
Infrastructure, Chips and Risks
The success of Saudi Arabia's AI program could depend on building out data center power supplies, switchgear and cooling. Nordic Semiconductor's new nRF71 system-on-chips combine Wi-Fi, Bluetooth LE, Thread and edge AI on a single chip. Supermicro is supplying cloud service providers with inference systems that combine NVIDIA GPUs and SambaNova SN50 chips.
Because of five vulnerabilities in the Reachy Mini robot, an attacker on the same network could take control without authentication. In other developments, SAP and Cohere are offering a sovereign AI platform for regulated industries, while NVIDIA is developing an open-source tool to identify the true causes of failures in distributed AI tasks.