pliny

A TypeScript meta-harness that treats LLMs as interchangeable brains, routes them through a mission loop, and insists on evidence instead of vibes.

A research-grade toolkit that treats safety guardrails as geometry, then uses SVD, projection, and feedback loops to excise them while trying to preserve the rest of the model.

How a mechanistic-interpretability project tries to identify, project out, and verify LLM refusal behavior without retraining.

A deterministic conlang system turns one numeric seed into grammar, vocabulary, and encoded agent-to-agent communication, revealing how linguistic design, obfuscation, and red-teaming collapse into the same machine.

A tiny static page turns innocent clicks into hidden behavior. The real story is not the verse, it is the gap between what the page shows and what the code does.

A client-side, Canvas-powered defense that keeps images local and turns plain HTML into a privacy tool.

It turns word submission into a ritual, then quietly keeps the ritual intact with local fallback, dataset snapshots, and a DataManager that refuses to let the experience die with the server.

The Android app maps plain language to hardware actions, then wraps the whole thing in risk checks, protocol translation, and live recovery.

It turns missing story inputs into a first pass, then keeps the draft moving.

ABRAXAS turns JSON prompts, shell output, and a self-copying loop into a closed system that feels less like an app and more like an experiment in autonomy.

A forensic archive of prompt loops, split personas, and safety refusals that look like exploits until you read the fine print.

A cartographic encyclopedia that turns abstract threat modeling into something you can explore, remember, and compare.

An open-source system that clusters noisy prompt scraps, verifies them by consensus, and turns a rumor mill into a real-time ledger.

A deep dive into the static browser toolkit that mutates language, hides payloads in Unicode, and reverse-engineers its own tricks.

By replacing fragile loops with matrix operations, this multi-modal toolkit makes data exfiltration invisible to modern EDRs, LLM filters, and social media compression.

How a pseudonymous researcher built a massive, open-source arsenal of linguistic exploits and glitch tokens to bypass AI safety guardrails.

How an open-source library of extracted system prompts reveals the hard-coded rules, hidden tools, and fragile guardrails governing commercial AI models.

A self-hostable research cockpit that perturbs prompts, races models in parallel, scores the contenders, and cleans up the result into a more controlled voice.