Offensive security
Hardware, RFID and SDR work on radio and telecom systems. iOS and Android apps taken apart with Frida and Objection. Red-team engagements and bug bounties, including recognition in NordVPN’s program.
Harish “Prin” Sharma · Founder & CEO, Cipher Academy
I break things, report them, and then build the fix. Hardware, machine-learning models, the networks that move information: I’ve been taking them apart since 2018. Now I run Cipher Academy, and I still ship my own products.
Move your cursor through the network. Every node you touch is compromised, and it spreads.
Every system trusts something: a radio signal, a training set, a crowd that seems to agree. My work is finding where that trust can be abused, then proving it.
Hardware, RFID and SDR work on radio and telecom systems. iOS and Android apps taken apart with Frida and Objection. Red-team engagements and bug bounties, including recognition in NordVPN’s program.
Data poisoning, model evasion and the hard limits of AI detection. My longest paper runs 97 pages on why entropy, burst, graph-neural, LLM-text and deepfake detectors fail.
I research how influence operations work once generative models are in the loop, and build the tools that catch them: coordinated-behavior detection, explainable disinformation scoring, OSINT monitoring desks.
Data poisoning, live
Each dot is a training sample from one of two classes. The line is the boundary a classifier learns from them, retrained in your browser as you scroll. Keep scrolling to slip in mislabeled samples, placed where they pull hardest.
Logistic regression, 600 clean samples, 400 held-out test points.
The other half of the job. Products and tools I designed and built, from live SaaS to research code.
Live · graphlit.co
Draw your architecture on a whiteboard. Graphlit turns it into a typed graph, has coding agents build the app against it, and checks on every commit that the code still matches the drawing. It is the production face of VIPL, my visual-intent programming research.
Live · sayit.cipher.academy
A pronunciation coach for English learners. Record a sentence and it scores you phoneme by phoneme, with IPA, tempo and pitch feedback. The same backend runs as a cloud SaaS or fully offline on a laptop. Moving to its own home at espeak.in.
Coming soon · iPhone
iOS app
The same coach, built for iPhone. Say a line and about five seconds later every word gets a verdict: Clear, Check or Unclear. Tap a word to see the sounds you aimed for next to the sounds the ear heard.
Rebuilding · inside StudyMe
Turns a STEM syllabus into a knowledge graph, then into narrated, animated lessons. The first video pipeline rendered over a hundred explainers before I scrapped it and wrote up why. The lesson engine is being rebuilt around authored vector scenes.
Working · defensive research
Paste one link. CoordFinder maps the accounts most likely coordinating around it and shows the evidence for each one.
An academy of my own, seven years leading security architecture, and freelance security work the whole way through.
2023Now
Cipher Academy · registered MSME
2018Now
Freelance · hardware, software, red team
20192026
Strat News Global
About 20 self-published papers from 2025–26 on AI-era information operations, adversarial ML and platform integrity, published through Cipher Academy. A selection is below. Copies on request.
AI-augmented coordinated inauthentic behavior and the collapse of algorithmic detection
Where entropy, burst, graph-neural, LLM-text, perceptual-hash and deepfake detectors break down; information-theoretic bounds on detection; an empirical review across 73 elections. A companion supplement covers the DSA, AI Act, FARA and GDPR.
Gap analysis across the DSA, AI Act and FARA: which behaviors each one reaches, and what falls between them.
How psychographic targeting works in AI-era influence operations, and where its real limits sit.
Data-poisoning attacks against machine-learning training pipelines. The demo above is the simplest case.
A survey of emerging detection methods and research programs against coordinated inauthentic behavior.
Operational impact, and the engagement paradox in AI-augmented influence operations.
NATO cognitive-warfare concepts, reflexive control and the “Three Warfares.”
A landscape survey of state and commercial influence-operations activity.
Contact / Delhi, IN
Need something broken, or built? Hardware to test, a model to harden, a product to ship, a team to train. Write to me.