Adam Guymon at Trustle describes experiments using Jev to assess cloud roles and select the least access needed for a request. Trustle sells identity and access management software; these are its own reported results, not an independently reproduced evaluation.

Break risk into narrower signals

  • A single broad risk question clustered Trustle’s staging assignments near the same score.
  • The team split the problem into role capabilities, privilege escalation, secret access and role type, leaving scope and usage calculations in code.
  • Supplying only the riskiest permissions made roles look uniformly dangerous. A broader access mix and typical permissions gave the model more context.
  • Trustle reports that fitting combination weights raised weighted agreement with a reference from 0.75 to 0.86 on held-out roles. That experiment used 110 public Google Cloud roles, not the staging assignments used for earlier iterations.

Route uncertain role selections

Trustle tested 66 access requests against answer keys it says were frozen before evaluation, with three runs per model. It reports 91.4% accuracy for Jev alone and 98.0% for Gemini 3.8-flash. A Jev-first policy that sent uncertain cases to Gemini reached 98.5% while sending 44% of decisions to Gemini.

Trustle reports a 0.1-second median and $0.10 per 1,000 decisions for Jev alone, versus 3.4 seconds and $3.08 for Gemini. These measurements belong to Trustle’s setup and were not reproduced here; the hybrid’s reported cost was $1.45 per 1,000.

Preserve the safety boundary

Trustle says Jev struggled with indirect privilege escalation, counting and requests needing multiple roles. Its fixes included focused questions, a none-of-these option and descriptive labels instead of numbers. A typed role choice prevents an invented option, but it does not make the chosen access appropriate; confidence thresholds need validation on the actual workload.

Caveat: Trustle’s small, internally authored evaluation and reference do not establish production security or generalize its accuracy and cost claims. No model calls or access decisions were independently tested here.