NeuralchemyI test AI agents by what they do, not what they say.
I'm Sanskar Jajoo — a self-taught AI security engineer. I build open-source tooling for prompt-injection detection and agent red-teaming: datasets & classifiers used 18k+ times on Hugging Face, and asrt-bench, which catches agent-safety regressions from real tool-call traces.
Tools that test and secure AI agents.
Practical, installable software — the kind you can run against a real system today.
asrt-bench
Fires poisoned inputs at a real agent and reads its tool-call trace for misuse — deterministic, no LLM judge. pip install asrt-bench
detonate ★ 10
Runs MCP servers & skills in a disposable, network-off container and probes them — reports what they actually do, not what the manifest claims.
reconRelate
A CLI-first tool that surfaces how reconnaissance findings connect to one another across a target.
Skills that make coding agents deliberate.
Installable skills for Claude Code, Codex and friends — making agentic AI work careful and verifiable.
socratic ★ 127
A self-questioning skill for coding agents — it interrogates a design across the domains that matter, then asks you only the decisions that genuinely need your call.
agent-structure ★ 18
A set of focused skills — authority, memory, verification, hands-on learning — each with a clear trigger and a clear exit condition.
Open prompt-injection data, used in the wild.
On Hugging Face — drawn on across the community to train and benchmark guardrails.
neuralchemy on Hugging Face
Binary and multi-axis datasets — intent, technique, severity — with group-aware, zero-leakage splits, plus a DistilBERT specialist series and DeBERTa detectors. 4 datasets · 9 models · 18k+ downloads.
Prompt-Injection Dataset
A binary injection / jailbreak corpus with group-aware, zero-leakage splits and deliberate hard negatives — benign prompts that look like attacks. A strong default for a detector.
prompt-injection-binary
Clean binary detection split, group-aware and leakage-free — the single-axis successor to the flagship corpus.
prompt-injection-intent
System extraction, role hijack, tool abuse, indirect injection, obfuscation and more — what the prompt is actually trying to do.
prompt-injection-technique
Encoding, payload splitting, context overflow, persona play, multilingual, few-shot poisoning — how the payload is constructed.
prompt-injection-severity
How dangerous the outcome would be if the injection landed — graded 1–3 for triage and ranking.
The ideas — and the code behind them.
Self-published papers on Zenodo, alongside the from-scratch implementations that back them.
AI In The Loop (AITL)
A systems taxonomy for closed-loop autonomous evaluation — the research foundation behind the ASRT line of work.
The Autonomous Sunk-Cost Fallacy
Stopping failures in agentic systems — why an AI doesn't know when it's actually finished.
The Modality Paradox
On the role of modality in autonomous LLM engineering, and where it quietly breaks.
Autopilot: Zero to Hero
ML and deep-learning models built from first principles — the code behind the AI-In-The-Loop research.
TinyMobileLLM
Running an 8-year-old phone — a Redmi K20 Pro, Snapdragon 855 — as an LLM host, to see how small models hold up on old mobile hardware.
Essays on autonomy & AI security.
Stuck in Its Own Loop
Why agentic systems fail to stop themselves — the autonomous sunk-cost fallacy, explained plainly.
AI Can't Satisfy Itself
The hardest problem in autonomous loops: teaching a model when it is genuinely done.
A Stock-Analysis AI Tool
A practical walkthrough — Python, the Gemini API, and FinBERT. The project that started all of this.
Teaching the hard parts, from scratch.
Course-style playlists that build the internals step by step — in Hindi and English.
Encoder from Scratch
Full playlist ↗
The stack I work in.
Let's talk about your model's blind spots.
Open to remote AI-security roles and freelance red-teaming / evaluation work. Based in Raipur, India — comfortable across time zones.