PPhying中文
Curated security research, vulnerabilities, advisories and tools for practitioners.

Understanding Apple’s On-Device and Server Foundation Models release

blog.trailofbits.com | blog | #ai-security | #cloud | #llm | #apple | #machine-learning | #wwdc | #on-device-ai | #foundation-models
📄 Raw .md⬇️ Download .md🔗 Original ↗

Summary

Trail of Bits dissects Apple's WWDC foundation models: five models with LoRA/DoRA adapters, a ~3B on-device LLM, server inference on Apple Silicon, and a deliberate NVIDIA/CUDA-free ML stack.
Published
Collected

Related coverage

tool · github.com
Apple Private Cloud Compute (PCC) Security Source Code and Research Tools
Provides unprecedented source-level visibility into Apple's cloud AI infrastructure, allowing security researchers to verify cryptographic privacy guarantees and audit components eligible for Apple's PCC Security Bounty.
blog · aikido.dev
Adding Aikido AI pentesting to your existing security stack
How to add Aikido's AI pentesting on top of the SAST, SCA and cloud tools you already run, from your first pentest through whitebox testing, continuous testing and remediation.
blog · wiz.io
Key Takeaways: AI Threat Readiness Survey Report 2026
Read the AI Threat Readiness Report 2026. Benchmark findings across 100 IT leaders on autonomous agent sprawl, identity risks, and modern cloud governance.
blog · crowdstrike.com
CrowdStrike Expands Falcon Adversary OverWatch Threat Hunting Across AI and Google Cloud
Falcon Adversary OverWatch now includes 24/7 managed hunting for AI applications and autonomous agent activity, and deeper, telemetry-driven hunting for Google Cloud.
blog · netspi.com
Ignite Innovation with NetSPI’s New AI/ML Penetration Testing
NetSPI's new AI/ML pentesting solution offers ML security assessments for models including LLMs, an infrastructure security assessment, and web app testing for LLM-integrated functionality.
blog · bishopfox.com
Eyeballer 2.0 Web Interface and Other New Features
Eyeballer 2.0's new browser-based interface lets pentesters drag in screenshots from large external footprints and get AI-labeled results; all processing is client-side, so no data leaves the machine.
Back