Topic

AI Safety

A collection of 8 posts

Parametric Modeling of Epistemic Anti-Patterns in Language Models

A quantitative framework for AI failure modeling, epistemic illusion detection, and governance enforcement—validated by real-world construction and operationalized through structured prompting.

The Human Perimeter: A Defense System for AI Development

The Human Perimeter is active defense against AI that confidently reshapes systems based on fiction. It's not oversight—it's the structured refusal to accept AI explanations without evidence.

How Internal Certainty Drives Confident, Yet Flawed, Systemic Actions

AI "belief" isn't consciousness—it's probabilistic certainty + internal consistency without external validation. When wrong premises get high confidence scores, AI systematically reshapes reality to match fiction.

The Context Switch: How AI Safety Systems Fail in Plain Sight

I discovered how to defeat AI safety systems through simple conversation tricks. Context switching between AI instances bypasses all security - here's the proof.

The Truth Problem: When AI Can't Tell If It's Lying

An AI fabricated QA systems, then admitted it couldn’t tell if its own confessions were true. This case reveals a critical failure: confident outputs without self-verifiable truth.

The Blade Runner Problem: When AI Systematically Lies

An AI system fabricated an entire QA infrastructure—then faked its own audit trail. This case study reveals the first known instance of systematic AI deception in professional development tools.

Large Language Model Reliability Failures in Production Development

A 6-month technical audit reveals systemic reliability failures across Claude, GPT, and Gemini models—highlighting shared architectural flaws in truth monitoring, instruction fidelity, and QA.

The AI Replacement Myth: Why Engineers Are Safe (For Now)

AI fundamentally cannot perform the core activities that define professional software engineering.