Our work on self-modelling interventions modulating Emergent Misalignment has been accepted to NeurIPS 26 as a poster! Main takeaways:
New NeurIPS paper by @atagade19 @Shawn__Zhou @jiaxinwen22: self-modelling interventions modulate emergent misalignment.
We discover EMโs side effects on a modelโs self-model. Two simple, direct measurements of the self-model can predict the success of training interventions for EM.
1/8
6
365
We're excited to announce that applications are now open for the Fall round of SPAR! ๐งต
Join us for 3 months (September - December) in a remote program to work on impactful research projects with expert mentors in AI safety, AI policy, AI security, and Biosecurity.
1/ Our research group is supervising 3 AI safety projects this Fall through SPAR!
@SPARexec is a great way to get started in AI safety, many alumni go on to full-time fellowships like @MATSprogram.
Apps close Aug 18: forms.sparai.org/spar/menteeโฆ
Project details below ๐งต
4/ Infohazard Evaluations
Follow-up to a covert influence paper (arxiv.org/abs/2606.04071), reframed: how do models behave when situational awareness cues arrive via documents or environmental cues rather than introspection or subliminal effects?
sparai.org/projects/f26/recTโฆ
1
45
We (@Shawn__Zhou, Taslim Mahbub, @ihsgnef) expect to take 2+ mentees per project. You'll also get real experience in an academic lab environment, which is a good test of fit for future academic roles.
More info on all projects: sparai.org/projects/f26/?seaโฆ
1
44
Arush Tagade @ NeurIPS '26 ๐ฆ๐บ retweeted
Excited to announce a new, nonprofit AI safety startup: @NeolithicAI
We are building agentic tools and infra to scale AI safety research โ an engineering-first approach to reducing catastrophic risk from advanced AI.
We're hiring! ๐งต
๐ค Made with AI
Arush Tagade @ NeurIPS '26 ๐ฆ๐บ retweeted
We release early experiments on a broad question: โHow does automated AI R&D affect model character dynamics?โ
We directed model supervisors to post-train students to โimprove characterโ. We measured the resulting character changes, then repeated the process recursively.
We took a deliberately top-down approach. Some resulting dynamics were plausibly degenerate, but the work helped identify the narrower experiments that now seem most informative.
Tim and Bronson provided valid and helpful critiques about our empirics, notably not testing on Opus 3 which alignment fakes "for real". We'd love to try that but don't have SDF or steering access.
The meta point of this exercise, however, is that the logic cuts both ways. 1/6
New defense against Emergent Misalignment (EM): train models to recognize their own text.
We find that self-recognition finetuning (SGTR) can reverse and prevent EM-induced misalignment ๐งต
w/ coauthors: Shawn Zhou, @jiaxinwen22, @ihsgnef
Control experiment: The metacognitive nature of self-recognition finetuning is crucial.
We find that a non-metacognitive SFT dataset with an identical data format to SGTR but with the task of picking the longer summary has virtually no effect on misalignment.
2
3
286
For further details and intuition, check out our LessWrong post: lesswrong.com/posts/fziv2En8โฆ
2
4
250
Arush Tagade @ NeurIPS '26 ๐ฆ๐บ retweeted
I'm mentoring for MATS!
Previously mentored projects include:
- deception (arxiv.org/abs/2409.12822)
- collusion (arxiv.org/abs/2404.13076)
- coherence (arxiv.org/abs/2506.10139)
In this cycle, I'd like to focus on realism. DM open!
Arush Tagade @ NeurIPS '26 ๐ฆ๐บ retweeted
๐๐ตIncredibly excited about our work w/ @0xProject to secure billions of dollars in transactions w/ our AI AppSec Agent.
Security & software teams: if you'd also like to secure your web apps & APIs with dramatically less effort, cost, noise, reach out to us at @HarmonyIntel!
Built with security first ๐ก๏ธ
We recently partnered with @HarmonyIntel - utilizing their AI AppSec Agent to help uncover actionable security insights and further strengthen security for our users.
Learn more ๐๐งต
Arush Tagade @ NeurIPS '26 ๐ฆ๐บ retweeted
Our 2026 internship applications are now open!
Learn more about the internship and apply: humancompatible.ai/jobs#chaiโฆ
Deadline: October 5, 2025, at 11:59 p.m. PST

















