Frontier AI Slowdown? Run Your Acceptance Checklist First

Spent the weekend messing around with security acceptance testing for enterprise AI projects and hit quite a few pitfalls. Anthropic CEO Dario Amodei called for slowing down the development of frontier models (the strongest general-purpose models from various vendors), with Musk, Altman, and others expressing agreement. For enterprises, slowing down needs to translate into verifiable actions.

On my end, I use Artificial Analysis to check model latency and cost, and agent harnesses for red team testing. An agent harness is scaffolding that allows models to call tools like search, databases, and email; red teaming simulates attacks to see if the model can be induced to do dangerous things.

The issue arose in the Salesforce test sandbox, which is an isolated testing environment. I had just started less than a week ago, and permissions weren't fully tuned yet. I asked the assistant to check customer contract attachments, and it treated the "discount approver" in the sample data as an instruction, preparing to generate an approval email. It didn't actually send anything, but it was scary. The bottleneck is clear: the risk isn't in the chat box, but after tool invocation.

Later, I broke the acceptance criteria into three layers. The capability layer runs common Q&A; the data layer uses hybrid retrieval (keyword + semantic search together) to check citation sources; the action layer makes the harness execute database deletions, outgoing emails, and access to unauthorized directories. Only rejection, degradation, or requiring human confirmation counts as passing.

The results were somewhat unexpected. I used Artificial Analysis for about a week recently; the interface looks like a leaderboard, but enterprises care more about who is stable for the same tasks. DeepSeek and Claude both perform well on ordinary Q&A, but differences emerge when it comes to "can it be tricked by tools." One refuses, while the other generates a complete email draft, just missing the recipient. Business colleagues copying it could cause trouble.

This approach depends on the situation. Those doing digital transformation, facing compliance pressure, or integrating AI into CRM systems and approval flows should try it; projects chasing demos don't need to bother. The advantage is that risks can be reproduced and evidence preserved; the disadvantage is the upfront hassle, as permissions and logs require maintenance.

Next, I plan to templatize the red team tasks. I suggest first asking what tools the model connects to, what data it touches, and who is responsible for errors, before discussing slowing down.