42 minutes ago
1
Your AI agent can write code, summarize legal briefs, and book flights. But can it do those things correctly 100 times in a row without accidentally leaking sensitive data or going rogue on a prompt injection? That’s the question Amazon thinks the entire industry is getting wrong.
Bryan Silverthorn, director of Amazon’s AGI Autonomy research lab, is set to present a new framework at VB Transform 2026 that reframes the conversation around AI deployment. The core argument: enterprises aren’t holding back on AI agents because the technology isn’t smart enough. They’re holding back because it isn’t predictable enough.
The capability-reliability gap
Silverthorn’s presentation, titled “Closing the capability-reliability gap,” takes aim at one of the AI industry’s favorite measuring sticks: benchmark scores. Those EVAL scores that labs love to trumpet on launch day? According to Amazon’s framework, they’re essentially vanity metrics for enterprise buyers.
Amazon’s new framework explicitly prioritizes consistency, robustness, predictability, and safety over raw performance on standardized tests.
An AI agent that correctly handles 95% of customer service inquiries sounds impressive until you realize the other 5% might involve unauthorized data access, hallucinated policy information, or actions taken without human approval. At enterprise scale, that 5% failure rate translates into thousands of problematic interactions per day.
Enterprises aren’t buying what labs are selling
A VentureBeat survey of senior technology leaders puts hard numbers behind the anxiety. Only 4% of respondents said they felt comfortable relying solely on model guardrails as a safety mechanism for AI systems.
The concerns break down along predictable lines. A full 40% of respondents flagged unauthorized access to tools and data as their primary worry when deploying AI agents. Another 27% pointed to prompt manipulation, the practice of tricking an AI into bypassing its instructions, as a top threat.
Amazon’s proposed solution involves decoupled systems and sandboxed environments. Rather than trusting the model itself to behave, you architect the surrounding infrastructure so that even a misbehaving model can’t cause real damage. Human oversight remains a non-negotiable part of the equation, particularly for any changes that affect production systems.
What this means for the broader AI market
VB Transform 2026, scheduled for July 14-15 in Menlo Park, will feature discussions on safe AI deployment strategies from multiple companies, including Waymo.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.
English (US) · 