So I guess everyone uses LLMs to think now
Reported by Hacker News1 source ↓
A security researcher says they do not use LLMs for published research because they enjoy doing the work themselves and cannot take credit for model output. They argue current systems still need skilled guidance and criticize how some companies publicize AI-assisted vulnerability findings.
Key points
- The author says all research published on their blog was done without AI
- The author argues LLMs need skilled researchers to guide vulnerability discovery
- The author criticizes public proof-of-concept videos and hashes that can create security fears
- A reported Tailscale privilege escalation required existing ACL access and Tailscale SSH enabled
Details
The author says they have never used AI for research published on their blog. They describe reverse engineering, coding and tracing source code as activities they do for enjoyment, and say they are uncomfortable publishing work done by an LLM as their own.
The author argues that LLMs can be powerful but still require skill and experience to direct, including in cybersecurity. They criticize what they see as AI companies presenting vulnerability discoveries as agent-led achievements, citing different accounts from Calif, OpenAI and Wiz about PHP vulnerabilities known as php2shell. The author says researchers and companies also publicize proof-of-concept videos and hashes in ways that can cause unnecessary alarm.
The article describes a Tailscale vulnerability that a researcher initially publicized with a hash and the phrase “privilege escalation.” According to the author, details posted the next day showed that exploitation required access under existing ACL rules and Tailscale SSH enabled on the affected machine, where the attacker could then gain root privileges. The author says the findings were published publicly; the article excerpt does not include its final account of the disclosure.
Written from coverage as of