Kushal Das on AI-assisted vulnerability research and defender burnout
A security maintainer's first-person account links AI-assisted vulnerability discovery to the human and infrastructure capacity needed to keep public-interest open-source systems defended, while making burnout and unequal access to capable models part of the security discussion.
Logged at IST: 2026-11-06 06:54 IST
What it is: Security researcher and open-source maintainer Kushal Das's first-person reflection on using LLMs to find vulnerabilities in infrastructure projects, and on the exhaustion and resource constraints that come with the work.
Gist: Das writes that since July he has spent long, tiring days probing open-source libraries, projects, and deployments, and says he had found more than 90 issues by the time of writing. He describes a July finding in which an open-weight model helped identify four zero-days in Java libraries, which he says were fixed, and helped chain them to bypass validation checks for digital signatures. These are the author's reported results, not an independently audited tally or technical disclosure.
The essay's broader point is about defender capacity. Much of the infrastructure he examines supports national research and education networks, maintained by small teams with limited resources. Das argues that powerful models, hardware, and people able to use them safely are becoming part of defensive capacity, while attackers can access similar capabilities cheaply. He also warns that the pace of vulnerability discovery risks burning out the people doing the work. His account is explicitly personal: he notes that he lacks the capability for full professional pentesting and does not provide the technical details of the findings in this post.
Source read: https://kushaldas.in/posts/i-am-tired.html
Newsletter angle: AI-assisted security is not only a question of model capability. The people, compute, permission, and sustained organizational capacity available to defenders, and whether that work is survivable, matter too.