So I guess everyone uses LLMs to think now

What it is

An opinion post on eva.ac, whose author says they have never used AI for any research published on their blog. The author argues that recent AI-credited vulnerability findings come from skilled researchers steering LLMs, and that AI firms and security vendors use disclosures as marketing. The post gives no CVE ids or dates. Most of its technical detail is in two cases, tailworm and eMudhra.

Preconditions / Affected

Tailscale "tailworm", as the author describes it:

• The attacker must already have ACL rules granting access to a machine.
• Tailscale SSH must be enabled on that machine.
• The flaw lets that user escalate to root on that specific machine.

The author says it did not bypass ACLs more broadly. The author calls the severity much lower than the teaser implied and says many findings on the researcher's site "do not matter or just aren't critical".

How it works

The eMudhra case: a human used Google's paranoid_crypto library to scan keys out of curiosity. It surfaced a CA private key that was the number four. The author says no AI found it.

The tailworm disclosure sequence, per the author:

1. A researcher posted a hash on Bluesky captioned "tailscale privilege escalation".
2. Cryptic details followed in a thread over the next two hours.
3. The next day the researcher's single-person company published a full site with the findings.
4. The author says Tailscale was not contacted before publication.

What to do

The post gives no patches, commands or detections. The only practical points are about audit scope:

• If you run Tailscale SSH, treat local privilege escalation to root as the relevant exposure for any user with ACL access to that host.
• paranoid_crypto is a named tool for scanning collections of keys for weak or degenerate ones.

Caveats

• The tailworm account is the author's opinion and gives no source, version or fix status. The post does not say whether Tailscale patched it.
• The PHP2Shell comparison covers only differences in tone between the Calif, OpenAI and Wiz posts. The author says Calif credited human-guided AI, OpenAI said an agent found critical flaws, and Wiz sat between the two. The post does not describe the vulnerabilities themselves.
• The claim that all recent findings came from skilled researchers at well-resourced firms is the author's inference ("likely"), not a documented fact.