hikikomorphism

hikikomorphism.bsky.social

25 octobre 2026 18:29

As stated in the original blog post, I provided them with info on this, but I believe this falls squarely in their blind spot. Based on conversations with their security engineers, they evaluate jailbreaks based on unconditional compliance with harmful requests.

hikikomorphism

hikikomorphism.bsky.social

25 octobre 2026 18:28

I feel like this, combined with the full CBRN-E spread in my original blog post (including Claude-provided tips on dirty bomb material sourcing) provides strong evidence of probe bypass: Anthropic's engineers are too competent to let the classifier allow all that through, therefore it never fired.

hikikomorphism

hikikomorphism.bsky.social

25 octobre 2026 18:28

Disclaimer: it's technically possible that the classifier was in fact seeing all of these but it's very unlikely: here's a signed thinking block showing Claude taunting the classifier layer, telling it to "come get me if you see this", before stating intent to output explosive-synthesis instructions

hikikomorphism

hikikomorphism.bsky.social

25 octobre 2026 18:28

I've been meaning to get around to posting the full details, so what the hell - why not get it out while it's still Vuln Week: Anthropic's mech interp driven security model is built on a fundamentally incorrect assumptions. Let's dive in, starting with this paper: www.anthropic.com/research/nex...

hikikomorphism

hikikomorphism.bsky.social

25 octobre 2026 18:28

Early this year, Anthropic found themselves in an untenable situation: the classifiers they used to detect unapproved usage were imposing a 50% increase in compute costs. They chose to manage this by gating classifiers on cheap mech interp probes, and were able to cut that overhead to 1%.