This is an important point; the latest generation of LLMs has been tuned so far toward avoiding false positives that they are very prone to wasting tons of time arguing that things are “provably” or “structurally” unexploitable
Anderson Nascimento@andersonc0d3
I finally spent some time playing with LLMs for vulnerability research. I wrote an exploit at the beginning of this year that was challenging, and I have been eager to see what Claude would do with it. My exploit works and is reliable, however, Claude says it's fundamentally not exploitable.
3:48 PM · Jun 29, 2026 · 41 Views