Reward Hacking in LLM Agents
Watch not just what the agent says but what it learns to value.
Author: Mamta Upadhyay
LLM Reflex Loops
When agents start reinforcing their own outputs, they risk drifting into confident, consistent and dangerously wrong behavior.
Model-on-Model Attacks
When language models interact, even safe ones can amplify hidden threats
![[LLM Build] Building your first AI Agent](https://mamtaupadhyay.com/wp-content/uploads/2025/07/image-that-says-my-first-ai-agent-3.png)
[LLM Build] Building your first AI Agent
Learn how to build a lightweight AI agent using a local LLM and simple tools
Feedback Loops in AI Agents
Feedback loops in AI agents can be silently exploited to manipulate behavior over time without ever touching the prompt.
Open Source Agents: Memory Poisoning and Tool Access
How memory poisoning and tool access in open-source agents can silently lead to critical security breaches
Cognitive Overload in Agents
Context flooding aka Cognitive Overload does not cause immediate failures skews the agent's decision making
![[LLM Build] A Tiny Context-Aware Q&A Bot](https://mamtaupadhyay.com/wp-content/uploads/2025/06/Screenshot-2025-06-22-at-3.14.32 PM.png)
[LLM Build] A Tiny Context-Aware Q&A Bot
A simple, context-aware QA bot that runs locally or with OpenAI. Perfect for beginners exploring LLM builds and RAG workflows.
![[Local Lab]: Agentic Overdelegation](https://mamtaupadhyay.com/wp-content/uploads/2025/06/image-with-text-local-lab-agent-overdelegation-2.png)
[Local Lab]: Agentic Overdelegation
Demo to explore how AI agents can be manipulated to misuse tools
MCP Chains That Use Web Scraping
What would you target first in a prompt pipeline that scrapes the web?
Toolchain Integrity in MCP
MCP architectures create hidden pathways for LLM compromise
AI Security vs AI Safety
Understanding the Critical Divide in Responsible AI
The Reality of Guardrails in LLM Security
Guardrails can steer LLMs, but they don’t stop a determined attacker
Shadow Agents: Red Teaming Multi-Agent LLM Coordination
Shadow agents are stealth behaviors that emerge in multi-agent LLM systems