Hot take: Most AI SOC vendors are selling a future that needs GPT-5 to work. We're stuck with GPT-4.

I've been trying to learn about AI SOC tools for my team's AWS setup. The marketing for a lot of these platforms sounds amazing—automated investigations, natural language queries, etc.

But when I actually test them, the demos often break on complex logic or need very specific phrasing. I'm using GPT-4 via API for some basic Terraform doc generation, and I see similar limits. For example, asking it to analyze a CloudTrail log for a specific IAM threat pattern works maybe 70% of the time.

```hcl
# This is simple, but real threats are messy
# Example: GPT-4 can flag this weird principal...
resource "aws_cloudtrail" "main" {
# ... config
}
# But can it reliably chain this with an unexpected S3 bucket policy change 5 minutes earlier?
```

Are the vendors overpromising? Is the current tech (GPT-4 class) actually good enough for real security ops, or are we all waiting for the next model jump?