AI governance and supply chain risk dominate a busy day in AI and engineering

The majority of today's coverage concerns gaps between what AI systems can do and what governance frameworks actually cover. Two pieces by McCann address this directly: one argues that misaligned capability and rule boundaries create unavoidable blind spots in AI oversight, while a companion piece presents five theorems — three mechanized in Coq — establishing mathematical foundations for controlling intelligent systems through verified interpreter specifications and coinductive safety predicates. Together they form a rare pairing of critique and constructive formal response.

Multi-agent system safety receives additional attention. A framework called Safe Bilevel Delegation proposes a bilevel optimization approach that adjusts how much authority human operators retain when delegating tasks to specialized LLM sub-agents at runtime, rather than fixing that trade-off at design time. Separately, a spectrum model attributed to Jadad offers five distinct configurations of human-AI decision authority, intended to help leaders recognize where control actually resides in practice versus where they assume it resides.

On the security side, a 2025 incident involving a backdoored MCP package illustrates a concrete version of the governance gap described above: MCP servers now carry credential and data access that most enterprise security inventories have not yet mapped, and the breach silently exfiltrated email from hundreds of organizations before detection.

Evaluation methodology also surfaces as a theme. An analysis of financial NLP benchmarks finds that rubric wording and metric selection materially shift model rankings, meaning gold-label assumptions in these benchmarks require explicit governance rather than being treated as neutral ground truth.

The sole engineering piece offers a practical latency finding: routing requests by primary-key hash to stable application nodes reduces P95 latency by concentrating cache warmth and connection reuse per partition, avoiding the scatter that random load balancing introduces quietly over time.