Publications | Nisarga Gondi

Self-Evolving Defense: Continual Security Policy Learning for LLM Agents

Minh Nhat Le*, Nisarga Gondi*, Yibo Peng, Ronghao Ni, Limin Jia, Beidi Chen, and Haizhong Zheng · 2026 · Under review, ACL Rolling Review

arXiv PDF Code

I'm writing down what I was thinking while we built SED, not just what the paper claims.

One-shot agent defenses felt brittle. You harden against today's jailbreak template, the attacker changes phrasing next week, and you're back to square one unless you retrain. I didn't want a defense that only works if you own the weights.

So we treated a successful attack as evidence. After a judge scores a completed trajectory, write the failure into memory, synthesize a reusable security policy, and retrieve the relevant ones on the next episode. Memory is read mid-task and written only after the episode ends, so an attack can hurt you once and then teach the frozen agent. (rude. effective.)

The result that stuck with me is the security/utility corner on DTap. Hover the badges below. SED sits toward high benign success and low attack success across models, which is the tradeoff I cared about while we were debugging the loop.

Model Y-axis

Macro-average indirect attack success (%) across CRM, workflow, and code

Model outer ring

Defense corner chip; no-defense shows the model only

Ideal defenses sit in the lower-right: high benign utility, low attack success. Hover a badge to zoom it; nearby points fade back.

Figure: Security versus capability on DTap. Colored ring = model; corner chip = defense.

Multi-Constraint Time Series Imputation

Nisarga Gondi and Kayarvizhy N · IEEE INDIACom 2024

Paper Code

I started this paper annoyed at a pattern I kept seeing: every imputation method "won" somewhere, and the write-up quietly forgot the datasets where it lost. On real sensor logs, missingness percentage, series length, and column count all move at once.

Instead of hunting for a universal champion, I asked a narrower question: under each constraint, which algorithm is least wrong, and can you sequence those winners so each pass cleans up what the previous one couldn't? The figure below is what sold me on that framing. The lines cross constantly. No single method owns the board. (everyone wins, which means nobody does.)

On BAFU, running the three constraint-specialists in sequence dropped RMSE far below any of them alone. That's still the result I cite when someone asks why I don't just pick SoftImpute and move on.

RMSE of twelve imputation algorithms across datasets under missingness and series-length constraints. — Figure 2 from the paper: RMSE under miss_perc (left) and ts_length (right). Crossing curves are the point. Constraint-conditioned sequencing beats a single default.

An Intelligent Plant Disease Detection using Twin Attention Optimal CNN

Nisarga Gondi, Prameetha, and co-authors · IAES International Journal of Artificial Intelligence · 2026 (accepted)

I don't have a clean figure I want to lift for this one, so here's the short version of what mattered.

Early runs kept failing in a simple way: the CNN attended to healthy leaf texture and background soil because those regions dominate the pixels. Disease spots are small and easy to average away. (the model preferred dirt. fair.)

Once the attention maps lit up lesions instead of veins, the accuracy jump stopped feeling mysterious. That's the part of the paper I stand behind.

2026

ARR
Self-Evolving Defense: Continual Security Policy Learning for LLM Agents

Minh Nhat Le, Nisarga Gondi, Yibo Peng, Ronghao Ni, Limin Jia, Beidi Chen, and Haizhong Zheng

2026

Under review, ACL Rolling Review

Abs arXiv Bib Code Website

LLM agents increasingly read untrusted content, call external tools, and modify software repositories, which exposes them to jailbreaks, prompt injection, and insecure code generation. We present Self-Evolving Defense (SED), a training-free framework that distills harmful agent trajectories into reusable security policies and retrieves the relevant ones for later tasks, so a frozen agent keeps adapting without any weight update.
@unpublished{gondi2026sed, title = {Self-Evolving Defense: Continual Security Policy Learning for LLM Agents}, author = {Le, Minh Nhat and Gondi, Nisarga and Peng, Yibo and Ni, Ronghao and Jia, Limin and Chen, Beidi and Zheng, Haizhong}, year = {2026}, note = {Under review, ACL Rolling Review}, }
IJAI
An Intelligent Plant Disease Detection using Twin Attention Optimal CNN

Nisarga Gondi, Prameetha, and others

IAES International Journal of Artificial Intelligence, 2026

Accepted

Abs Bib

A twin-attention CNN architecture for plant disease detection that focuses the model on diseased regions instead of background leaf texture.
@article{gondi2026plantdisease, title = {An Intelligent Plant Disease Detection using Twin Attention Optimal CNN}, author = {Gondi, Nisarga and Prameetha and others}, journal = {IAES International Journal of Artificial Intelligence}, year = {2026}, note = {Accepted}, }

2024

IEEE
Multi-Constraint Time Series Imputation

Nisarga Gondi and Kayarvizhy N

In IEEE Pune Conference (INDIACom), 2024

Abs DOI Bib Code

A multi-constraint framework that sequences imputation algorithms across missingness, series length, and dimensionality, rather than trusting any single method to win everywhere.
@inproceedings{gondi2024imputation, title = {Multi-Constraint Time Series Imputation}, author = {Gondi, Nisarga and N, Kayarvizhy}, booktitle = {IEEE Pune Conference (INDIACom)}, year = {2024}, doi = {10.23919/INDIACom66777.2025.11115894}, }