AI Engineering Signal #73
Alibaba bans Claude Code internally over alleged ANTHROPIC_BASE_URL backdoor
Signals
Alibaba bans Claude Code internally over alleged ANTHROPIC_BASE_URL backdoor
audit any Claude Code deployment using custom base URL routing before next release cycle.
Reuters
Zuckerberg: AI agents behind internal expectations
agent-dependent roadmaps and procurement timelines need reassessment now.
TechCrunch
DeepSeek V4 Flash runs 1M-token context locally on RTX 5090
long-context workloads that required cloud APIs have a viable local inference path.
BPE tokenization creates exploitable alignment gaps in LLMs
add token-boundary adversarial cases to red-team checklists for safety filters.
ArXiv
Anthropic in custom chip talks with Samsung
a closed deal shifts inference supply chain away from NVIDIA; flag for procurement planning.
TechCrunch
Energy Dept. directs data centers onto backup generators during heat wave
operators without tested generator failover now face on-demand grid-relief compliance pressure.
Web
The Take
The Alibaba ban and the BPE alignment paper arrive together: AI tooling trust is fracturing at the infrastructure layer. Enterprises that haven't audited custom endpoints and token-level inputs are exposed on their largest attack surface.
Subscribe
Related Signals