The alpha, twice a week
That was 2 min.
Get the best new moments in your inbox every Wednesday and Saturday. It's free.
Check your inbox
We sent a link to your address. Click it to finish signing up. Not there in a minute? Look in spam.
or
MTS · 2 min · Oct 17, 2026
Customers can define guardrails for their own deployments
Devon Thomas
CEO, Abliteration
Theo JaffeeHost · Co-founder & host, MTS
One of the first members of a16z's new media team; earlier hosted the Emergent Ventures-backed Theo Jaffee Podcast.
Sophia DewHost · Host, MTS
About this segment
Thomas says the company plans to focus on models for customer use cases, not the application layer. Customers can set organization- or project-level rules, while some refusals remain hard-coded.
From “Model Refusals Don't Stop Hackers | Devon Thomas”, Oct 17, 2026. Watch the full episode
Thanks. We'll look at and fix what's off.
“we do have a feature where in like natural language or otherwise they can set their own guardrails that works for like their organization”
Takeaways
- The business aims to provide models for customer use cases.
- Customers can customize guardrails at the organization or project level.
- Some categories, including CSAM and self-harm, still trigger refusals.
In this episode
Full episode- 0:45Cyber defenders welcomed access to less-restricted models2:29
- 3:05Defenders need to test systems before attackers do3:12
18:06Customers can define guardrails for their own deployments2:12
- One more moment in the full episode
Transcript
Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.
- 18:12
doing things it's hard to know people's motives and and agendas and what uh who they're trying to appease i guess right so what's the what's the business model in the long term is it going to be like an inference provider like a platform for red teaming or like a contractor for uh like the military or other people who need to do cyber stuff yeah kind of that that last term so like are always our goal is going to be like to have the best models for the use cases that matter most to the customers that we have and we have quite a few customers and like like i said from enterprises to small business uh small startups and so um like the goal will always be to provide the best models for their use case and we don't ever have like plans to move more into the application layer we'll let the our customers kind of handle that so where the harness layer is what'll be the like
- 19:01
our goal will have the best models for those those use cases that our customers care about do you obliterate on bio refusals too or just cyber so uh no we obliterate on refusals in general and then we do have some hard-coded we do actually still have refusals in the model if you ask some of our customers they've hit some of them but uh wow yeah yeah but uh but uh not yeah so yeah it's not just like a uh and then like I always say like our innovation on it is to try to like let our customers set their own guardrails and that's like something we don't really get much press for but we do have a feature where in like natural language or otherwise they can set their own guardrails that works for like their organization and uh so yeah so and then you can let's say like maybe you do need looser uh cyber guardrails um or and then you may want
- 19:46
more stringent NSFW guardrails or maybe you want agentic guardrails like hey don't ever go truncate a database or don't uh do something else so like the idea is like more of an inversion of control right you're saying okay you as an organization can kind of divide Find the guardrails for your organization. You can set that at an organization level or a project level as well. What are some of these hard refusals, though?
- 20:09
So CSAM, we have hard refusals. Also, like self-harm, we have hard refusals as well. So, yeah.