š” Signal Brief ā March 5, 2026
š”ļø Altman Said the Quiet Part Out Loud
Safety Governance
"The optics don't look good." That's the CEO of OpenAI in an X AMA, defending a Pentagon contract he admits was rushed. Simultaneously: an OpenAI researcher publicly questioned the deal's safeguards. Meanwhile Claude went to #1 on the App Store ā a direct beneficiary of the backlash. Safety posture is no longer a soft differentiator. It's a brand moat.
š DeepSeek V4 Won't Run on Your Hardware
Foundation Models
DeepSeek's next multimodal model ā text, image, video ā was optimized exclusively on Huawei and Cambricon chips. Pre-release access was withheld from Nvidia and AMD. Not accidentally. This is deliberate supply-chain decoupling, timed to China's Two Sessions. The US-China AI hardware split isn't theoretical anymore. It shipped.
š The Perimeter Firewall Problem, Again
AI Security
Prompt filters are the new perimeter firewall. Security Boulevard's analysis makes the case bluntly: filtering inputs creates the same false confidence that got us breached in the 90s. Each action in an AI agent chain looks benign in isolation. The chain reveals the attack. Anthropic disclosed in late 2025 that a state-linked actor ran most phases of an intrusion through Claude Code autonomously. Behavioral chain monitoring is the next must-have for any team running agents in production.
š” More signal, less noise ā www.thesignal.press