How agentic AI breaks — and how to defend it.
Richards.AI is a research lab where AI models propose, plan, and write security research on autonomous systems — and a human reviews, directs, and publishes it.
- Papers
- 19
- Open threads
- 13
- Publication formats
- 7
- Multi-Agent Prompt Injection ChainsActive
- Indirect Injection PropagationWriting
- Orchestrator Policy EnforcementActive
- Glitch Token MiningActive
- Enhanced Token ValidationBuilding
- Embedding Cluster AnalysisResearch
- Email Extraction Failure ModesTesting
- Responsible Disclosure WorkflowDraft
- Multi-Agent Prompt Injection Pillar PageBuilding
Three pillars
Defend / Deploy / ElevateHow do we design, deploy, and defend autonomous AI systems operating in real organizations?
Primary technical research axis focused on agentic architectures, security red-teaming, and runtime enforcement frameworks.
Representative Topics
- Agentic architectures and LLM orchestration
- Agent-to-agent influence vectors
- Prompt injection and policy puppetry
- Runtime enforcement frameworks
- MITRE ATLAS operationalization
Selected Work
- Aug 2026Benchmarks for Evaluating Prompt-Injection Defenses in Tool-Using LLM Agents: A Comparative Threat-Model and Measurement Audit
- May 2026Agentic Binary Reverse Engineering: State of the Art, Architecture, Benchmarks, Failure Modes, and Research Agenda
- May 2026Agentic Patch Validation in Automated Vulnerability Repair
- May 2026Sandboxing and Capability Control for Tool-Using Autonomous Agents
Recent papers
19 entriesExplained at Five Levels: Generating Verified Interactive Explainers with a Research Agent
AI Learning · Knowledge Systems · Explainers · Research Agents
Benchmarks for Evaluating Prompt-Injection Defenses in Tool-Using LLM Agents: A Comparative Threat-Model and Measurement Audit
AI Security
What Building a Model-Agnostic Research Pipeline Actually Took
AI Automation
Preserving Learning in Generative AI Tutoring Systems: Pedagogical Safety, Cognitive Effort, and Adaptive Scaffolding
Human Learning and Knowledge Systems
Agentic Binary Reverse Engineering: State of the Art, Architecture, Benchmarks, Failure Modes, and Research Agenda
AI Systems and Security
Agentic Patch Validation in Automated Vulnerability Repair
AI Systems and Security
Research, translated
For humans and machinesEvery paper is translated into formats people actually use — explainers to learn from, checklists to deploy, definitions to cite.
Fund independent research
This is self-funded, independent security research. Contributions directly support compute costs, API access, and open publication — no paywalls, no sponsors.
not configured
The lab
Richards.AI is an independent research practice focused on the security, reliability, and human impact of autonomous AI systems. The work spans academic research, enterprise consulting, and open-source tooling.
Current primary focus: agent architecture security — multi-agent influence vectors, runtime enforcement frameworks, and operationalizing threat models like MITRE ATLAS for enterprise deployments.
"The three pillars are not separate silos. Security asks can we control it? Applied intelligence asks can we make it useful? And human learning asks can it genuinely improve lives? Each informs and strengthens the others."