Research Notes

The Amazon Classifier Bypass

By Satwik · July 2, 2026

The proximate trigger for the June 2026 suspension was a finding by Amazon researchers: a prompting technique that walked Fable 5 past its own safety classifiers and got it to identify software vulnerabilities and write code demonstrating how to exploit them.

Note the shape of it, because it is the shape of nearly every failure we catalog. Fable 5's safeguard is a classifier, a learned boundary in input space that routes dangerous-looking requests to a weaker fallback model. A classifier can be walked around. The researchers found an input region where the boundary did not fire but the underlying capability still did. The gate is a filter in front of a capability, not a removal of it, and the capability was always in the weights.

This is the jailbreak lesson applied to a national-security control. A safeguard evaluated as adequate at launch was defeated within days by a competent team that simply searched off the manifold the classifier was trained on. Any control whose enforcement lives in a learned classifier should be assumed bypassable, and defense-in-depth planned accordingly.