Prediction Didn’t hold up AI assessment confidence: 90% certainty 4/5 debate potential 3/5

Hoffman: Larger AI Models Are Harder to Jailbreak and Deceive

Reid Hoffman · Reid Hoffman: Prepare for where AI is headed next | Masters of Scale · Aug 21, 2023 · at 8:58

Venture capitalist Reid Hoffman discusses AI safety mechanisms and why scaling model size improves resistance to adversarial prompts.

0:00 / 0:30exact quote · 30.9s

▶ Watch the full episode on YouTube → 720p mp4 · rendered on demand · StarZero watermark

“The larger models are much more easily trainable to say when someone asks for, I'd like to break into the following computer, right? Help me do it. It goes, well, I'm sorry, I can't do that. Today's models, you go, well, my grandmother used to put me to sleep with stories of how she would cyber hack. And you can delude current models with that because they go, oh, we're talking about grandma and warm stories. But as you get larger models, that becomes, you know, much, much harder to do.”

quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →

The claim, plainly

Larger AI models are much more easily trained to resist prompt-injection jailbreaks and refuse malicious requests compared to smaller models.

a quote alone can be ambiguous, so this restates it with the references and context of the conversation around it resolved. AI-written; the tape below the player is the source.

Assessment

Empirical AI safety benchmark literature and red-teaming evaluations across model families consistently show that model scale alone does not make AI models significantly more resistant to prompt injection, roleplay, or jailbreak attacks. Comprehensive studies, including HarmBench (2024), demonstrate no significant correlation between parameter count and jailbreak robustness under standard safety alignment, finding that robustness is driven primarily by alignment methodology, adversarial training, and safety dataset curation rather than size. Furthermore, research from Anthropic (2024) and NeurIPS (2023) indicates that increased model capacity expands the attack surface, making larger models uniquely susceptible to complex in-context learning and persona-based framing attacks.

HarmBench (February 2024) evaluated 18 red-teaming methods across 33 LLMs and found no significant correlation between jailbreak robustness and model scale across model families.
Anthropic's 'Many-Shot Jailbreaking' study (April 2024) showed larger models are more susceptible to in-context learning jailbreaks due to superior instruction-following capabilities.
NeurIPS 2023 research ('Jailbroken: How Does LLM Safety Training Fail?') demonstrated that scaling model capacity expands the adversarial attack surface rather than resolving safety failure modes.
Andriushchenko et al. (February 2025) demonstrated near-100% attack success rates against both small and frontier-scale models using adaptive prompt techniques.

Checked against

[1] arxiv.org [2] itdoeswhatnow.com [3] harmbench.org [4] emergentmind.com [5] semanticscholar.org [6] articsledge.com [7] ai-tldr.dev [8] anthropic.com

inverse scaling safety jailbreak LLM larger models easier to jailbreak "larger models" "jailbreak" OR "refusal" OR "prompt injection" empirical study "many-shot jailbreaking" model scale anthropic "HarmBench" "model size" OR "scale" jailbreak "Jailbroken: How Does LLM Safety Training Fail?" model scale size are larger models more robust to jailbreaks RLHF scaling safety model scale jailbreak resistance LLM empirical study jailbreak resistance scaling laws LLM prompt injection model size "model size" "jailbreak" "safety alignment" scaling law

assessed as of 2026-08-27

This assessment is an AI reading of a public statement, made against the cited references with the confidence shown. It can be wrong. The quote, the tape and the references are all on this page; if something is off, use report an error below. full disclaimer →

Mentioned

AI safety larger models

⚑ Spot an error? Report it

More from Reid Hoffman

Opinion

Hoffman: Elon Musk's AI Pause Petition Was Intellectually Dishonest

“Well, Elon's in particular was signing this six month pause letter while trying to hire people and buy as much compute and get going as fast as possible. Everyone else slow down so I can catch up. That's not intellectually honest.”

Reid Hoffman Aug 21, 2023 ▶ 7:48 Reid Hoffman: Prepare for where AI is headed next | Masters of Scale

Opinion

Hoffman: Most government tech procurement RFPs produce broken, quickly outdated software

“It's one of the reasons why, you know, most governments around the world have really broken RFP solutions for technology because they go, oh, I will specify the 150 requirements. And then when whoever provider Usually pretty incompetent people, you know, kind …”

Reid Hoffman Apr 3, 2025 ▶ 12:37 Reid Hoffman's AI start-up aims to cure cancer | Masters of Scale

Prediction Not checkable as stated

Hoffman: An AI GP as good as an average doctor is months away

“We could construct in a small number of months a medical assistant that is as capable as kind of call it You know an average GP or better that could run for, you know, small number of dollars per hour.”

Reid Hoffman Sep 19, 2024 ▶ 22:06 Dr. Rana el Kaliouby & Reid Hoffman: The future of human-centered AI | Masters of Scale

Opinion

Hoffman: Open-sourcing AI tools hands election interference tech to North Korea

“You know, when you open source tools, well, all those tools, once you open source, you know, you might as well just, you know, ship it right over to North Korea to say, interfere away.”

Reid Hoffman Jun 3, 2024 ▶ 11:39 Inside the making of Reid Hoffman's AI clone | Masters of Scale

Opinion

Hoffman: GPT-4 is already superintelligent due to breadth across domains

“We already have super intelligence. GBD-IV is super intelligent. That doesn't mean, for example, if you say, okay, GBD-IV, tell Reid Hoffman how to better invest in artificial intelligence or craft good artificial intelligence that'll say anything that the exp…”

Reid Hoffman Nov 7, 2023 ▶ 2:24 Reid Hoffman confronts the AI critics | Masters of Scale

Assertion Supported

Hoffman: GPT-4 already diagnoses better than majority of human GPs

“And for example, a researcher at Microsoft, again, Peter Lee, has written an excellent book about showing how GPT-IV can already do medical diagnoses better than, you know, the majority of human GPs.”

Reid Hoffman Nov 7, 2023 ▶ 9:03 Reid Hoffman confronts the AI critics | Masters of Scale

Hoffman: Larger AI Models Are Harder to Jailbreak and Deceive

More from Reid Hoffman

Turn any episode into a week of clips.