Joshua Saxe

What it was like working on LLMs and security at Meta (2022-2026)

I loved my time at Meta, and I also counted the days between equity vests and daydreamed about quitting on the morning after almost every one.

May 20 • Joshua Saxe

95

6

4

Against technocentrism in AI safety

Much of AI safety operates from the technocentric view that the central object of analysis, and the central source of risk, is the model itself.

May 12 • Joshua Saxe

13

5

2

April 2026

Exploits don't cause cyberattacks

On thinking clearly about frontier AI advances and cyber conflict

Apr 9 • Joshua Saxe

17

7

5

March 2026

What to ask vendors at RSA

Hard questions to ask vendors selling AI for security and security for AI solutions

Mar 23 • Joshua Saxe

15

3

3

Dual-use AI security tools advantage attackers

The current trajectory, in which the most effective AI tools focus on dual use vulnerability discovery, and defensive tools lag behind, advantages…

Mar 20 • Joshua Saxe

13

1

3

January 2026

How to defend an exploding AI attack surface when the attackers haven't shown up (yet)

The dilemma in AI agent security in 2026 is that organizational attack surface is expanding at comic pace but, at least as of January, the attackers…

Jan 6 • Joshua Saxe

34

6

On apples, oranges, and classical ML versus LLM security performance

(Meta: I struggled to write this post engagingly, so am expressing my points in dry prose, below, and in Claude-authored Michael Lewis potboiler prose…

Jan 2 • Joshua Saxe

16

1

1

December 2025

Swallowing the reality pill in AI agent security

And why we need to focus on low friction security technology in the era of AI acceleration

Dec 23, 2025 • Joshua Saxe

21

1

4

October 2025

What Should Practical AI Security Folks Think of 'If Anyone Builds It, Everyone Dies'

(Where “it” is artificial superintelligence, “everyone” is the human race, and the policy prescription is to halt or heavily regulate AI development.)

Oct 24, 2025 • Joshua Saxe

10

4

A take on how we should approach securing AI agents now and in the long term

Presentation I gave at The Curve, and as a keynote at the AI Security Forum

Oct 8, 2025 • Joshua Saxe

6

September 2025

AI security notes 9/15: We can get control of prompt injection without any technical miracles

Prompt injection often gets talked about as though it will require deep machine learning breakthroughs to overcome.

Sep 16, 2025 • Joshua Saxe

9

3

2

August 2025

AI security notes 8/13/25: GPT-5, the new incremental capability / exponential adoption regime, and what it means for AI security

TL;DR GPT-5 signals a shift from a ‘revolutionary regime’ (2019-2025) in which major AI releases unboxed significant technological surprises, to an…

Aug 13, 2025 • Joshua Saxe

3

3