AI model releases and agent security #127
Today's Letter
- Anthropic, Claude Fable 5.1 and Mythos 5.1
- Mistral, data-training opt-out controls documented
- NVIDIA, CrowdStrike announce SafeMind security system
Anthropic, Claude Fable 5.1 and Mythos 5.1
- Anthropic introduced Claude Fable 5.1 and Claude Mythos 5.1 for coding, knowledge work, and scientific research
- Both models share the same underlying system, with different safeguards and access policies
- Fable 5.1 is generally available; Mythos 5.1 is restricted to trusted access programs
- Anthropic reports 25% lower typical token-billed costs and up to approximately 45% lower costs for highly agentic workloads
- Cybersecurity safeguards produce 60% fewer false positives and support vulnerability discovery without exploit development
- Enterprise Frontier Safeguards will provide customer-controlled storage and zero-data-retention privacy in phases later this fall
- Mythos 5.1 biology access is being developed with the US government, with scientist enrollment expected soon
Source: anthropic.com
Mistral, data-training opt-out controls documented
- Mistral may include user inputs and outputs, including conversations and documents, in model training programs
- Vibe users are not opted out by default and can disable training in settings
- Vibe Enterprise customers are opted out by default, with opt-in controlled by administrators
- Mistral Studio and API customers can disable anonymous improvement data in the Admin panel
- Vibe and API opt-out controls are separate and must be configured independently
Source: help.mistral.ai
NVIDIA, CrowdStrike announce SafeMind security system
- NVIDIA and CrowdStrike announced SafeMind, an agentic cybersecurity system built into the CrowdStrike Falcon platform
- SafeMind combines CrowdStrike cyber models and agent harnesses with NVIDIA Nemotron open models
- Nemotron 3 Ultra orchestrates the defensive harness, while a fine-tuned Nemotron 3 Super powers rule generation
- CrowdStrike’s internal evaluations reported 99% lower cost and higher accuracy for the Blue Solano model than leading frontier models
- Red-team and blue-team agents continuously test exploits, generate detections, validate findings, and promote protections
- CrowdStrike also introduced Falcon IQ, an agentic workforce of more than 50 agents for assessment, prioritization, and remediation workflows
Source: blogs.nvidia.com
Jocoletter curates AI, software, and product trends for developers and builders.
#Anthropic #Mistral #NVIDIA