DetectThe company
Is your AI SOC agent still right?
AI agents can now close your alerts, tune your detections, and decide what gets escalated. DetectTrace continuously checks their work, on your cases, so a quiet model swap or a bad prompt update does not become a missed incident.
The thesis
An AI SOC agent can close 95% of your alerts correctly and you still would not know it, because nobody is checking whether the 5% it got wrong is getting worse.
The product
The trust layer your AI SOC needs.
DetectTrace sits above whatever AI SOC agent you already run, vendor-built or homegrown. It replays the agent against your real cases, checks whether it gathered the right evidence and reached a defensible verdict, and keeps watching after deployment so you know the moment quality drifts.
Status: the first agent evaluations are currently being built. The product follows the cases, not the other way around.
The free CLI
detect-forge
Open source. Runs locally. Tells you which of your detection rules still catch what they claim. A rule that silently stopped firing looks identical to one that never fires at all, until something checks. This is the layer below agent trust.
pip install detect-forgeThe standard
Four questions, both directions
Can I validate this claim? Can it even be validated? Where is the evidence trail? Is it still true after the model, prompt, or rule changed? Vendor accuracy claims and agent verdicts get graded against them. So does everything DetectTrace ships.
The build log
Get the build log.
Quantitative AI Security. One issue a week: what shipped, which agent evaluation broke, what it cost, who paid, and the numbers, including the ones that hurt. No hype, no growth hacks. Unsubscribe in one click.
Or read it on the web first: james-bower.kit.com/quant_newsletter