Critics Slam EA Alignment Camp Over Security Engineering, Sparking AI Safety Route Debate

A debate over AI safety approaches escalated from September 20 to 22: critics accused the EA/doomer camp of ignoring cybersecurity, while supporters pushed back with public materials from top AI safety organizations to rebut the "inaction" charge.

Confirmed

Zack Korman posted a critique of the EA/AI doomer camp, arguing they believe security is impossible to achieve and that only alignment can save humanity—but in reality they don't understand security engineering; the theory just makes them feel good, and this dangerous phenomenon should end (m1).
Gary Marcus shared the post in support, adding that if EAs and doomers truly cared about AI safety, they should study cybersecurity and use every means to control AI (m2).
trevposts rebutted the criticism, noting that cybersecurity-style AI safety has long been the main route of leading AI safety organizations, citing Our Approach to AI Safety and Security from OpenAI, which he described as the biggest funder in the space (m3, m4).
In m4, trevposts quoted the document, listing valuable research directions including detecting or mitigating harmful AI behavior via harmful-output classifiers, control protocols, and various interpretability techniques.
panickssery reshared in support, saying model theft prevention and constraining out-of-control AI behavior are exactly the focus areas where top organizations keep investing (m5).

Why it matters

The dispute exposes a route divide within the AI safety community: alignment research versus cybersecurity engineering, with opposing views on who is actually reducing risk.
By citing OpenAI's official safety roadmap as evidence, trevposts pulled the argument from emotional accusations back to verifiable facts, offering a reference for understanding what mainstream AI safety organizations actually work on. Whether either side responds further deserves continued attention.

2026-09-20 ~ 2026-09-22 · 5 related posts

Primary sources

Cybersecurity Has Been Mainstream AI Safety for Years, Researcher Counters — trevposts · 2026-09-21 23:45 (UTC+08:00)
Receipts: OpenAI's Safety Approach Paper Details Security-Based Mitigations — trevposts · 2026-09-21 23:55 (UTC+08:00)
AI safety organizations have long worked on cybersecurity and misuse, push back on critics — panickssery · 2026-09-22 01:06 (UTC+08:00)

Critics Blast EA Aligment Crowd for Claiming Security Is Impossible Without Understanding It — kevinnbass · 2026-09-20
Gary Marcus: EA doomers should embrace cybersecurity instead of vague alignment talk — AlexTensor · 2026-09-21
[source] Cybersecurity Has Been Mainstream AI Safety for Years, Researcher Counters — trevposts · 2026-09-21
[source] Receipts: OpenAI's Safety Approach Paper Details Security-Based Mitigations — trevposts · 2026-09-21
[source] AI safety organizations have long worked on cybersecurity and misuse, push back on critics — panickssery · 2026-09-22