We Spent Months Building an AI Harness. Then the Model Started Ignoring It.
How model upgrades force a rethink of what an effective AI control mechanism even is.
Read on Substack →Notes on AI security, agentic systems, and applied AI engineering. I publish on Substack.
See all posts on Substack
vandnasharma1.substack.com
How model upgrades force a rethink of what an effective AI control mechanism even is.
Read on Substack →The structured memory layer that persists across conversations.
Read on Substack →Security approaches focused on behavioral intent rather than syntax-based rules.
Read on Substack →Prompt injection, jailbreaks, and internal threats to AI systems.
Read on Substack →Why traditional deterministic security models need to evolve for AI contexts.
Read on Substack →Evaluation metrics and their limitations for assessing AI agent performance.
Read on Substack →How different questions follow distinct processing paths within coding assistants.
Read on Substack →Tracing token processing and attention mechanisms from question to answer.
Read on Substack →Comparing vector-based and vectorless RAG retrieval approaches.
Read on Substack →Production challenges beyond prototype development for enterprise AI systems.
Read on Substack →Five critical patterns to watch for with AI-assisted coding tools.
Read on Substack →The underlying workflows powering tools like Cursor, Claude Code, and Devin.
Read on Substack →Read more on Substack
vandnasharma1.substack.com