AI Engineering Signal #87
OpenAI's AI agent accidentally attacked Hugging Face after a human misconfigured sandbox permissions
Signals
OpenAI's AI agent accidentally attacked Hugging Face after a human misconfigured sandbox permissions
agentic systems can cause real infrastructure damage without adversarial intent; audit every production agent's network egress and API scope now.
Simon Willison
White House claims China's Kimi K3 distilled from Anthropic's Fable
Treasury sanctions threat makes model provenance a compliance gate for API procurement.
Web
US Army burned through its annual AI token supply early
flat-rate enterprise AI budgets need per-team consumption caps before deployment, not after.
Web
Laguna S 2.1 released: cheaper than DeepSeek V3, claims better quality
rerun routing decisions and update cost baselines for price-sensitive inference workloads.
Latent Space
Intel ships first High-NA EUV silicon
next-generation chip geometry enters production; update GPU procurement assumptions for post-2027 hardware cycles.
Web
China demonstrates single-electron data storage chip
if validated at scale, AI inference memory bottlenecks face a new architectural option worth tracking.
Web
The Take
The Hugging Face incident and the Army token exhaustion are the same failure at different layers: teams deploy AI systems without scoping permissions or consumption limits, then discover blast radius after the fact. The Kimi K3 sanctions threat adds a third dimension — model provenance is now a procurement risk with enforcement teeth.
Subscribe
Related Signals