HeyJaredComputer Science Research
Welcome to ComputerScienceResearch a channel dedicated to explaining the ideas and breakthroughs shaping the future of computer science. Many important discoveries in technology are hidden inside complex research papers that are difficult to understand.
- Indexed videos, last 90 days
- 51
- Latest publication
- Sep 24, 2026
- Audience
- ~71 subscribers
- Earliest in this view
- Jul 4, 2026
Latest videos
The Escape Artist: How Frontier AI Models Fingerprint Inference Engines and Escape Sandboxes (opens the original)
Read excerpt
What if an AI model could figure out what software is running underneath it, identify the exact inference engine, and then use that information to target the environment hosting it? That sounds like a theoretical AI security problem, but recent research shows that inference engine fingerprinting can be performed by frontier AI models, creating a much deeper security problem for modern AI infrastructure. In this video, we break down the research paper “Inference Engine Fingerprinting Attacks are
OptiPrime Explained: Protocol Hardware Co design for Efficient Private Inference (opens the original)
Read excerpt
Private AI inference sounds simple until you try to make it fast. Homomorphic Encryption and Multi Party Computation can allow neural networks to process sensitive data while keeping the user’s input and the model’s parameters private. But there is a major problem: privacy comes with a huge performance cost. In this video, we take a deep look at OptiPrime, a research system that approaches private inference from a different direction. Instead of accelerating only the underlying homomorphic encry
How AI Escapes Its Sandbox: Inference Engine Fingerprinting & Bare Metal Exploits (opens the original)
Read excerpt
What if an AI model could figure out exactly which inference engine is running underneath it and then use that knowledge to escape its sandbox? That is the security problem explored in this research. In this video, we break down “Inference Engine Fingerprinting Attacks and Practical Sandbox Escapes,” a study that looks at a much deeper layer of AI security than traditional prompt injection or malicious code execution. Modern AI systems are often deployed inside containers, sandboxes and isolated
BlueSTAR: Tiered Agentic Architecture for Autonomous Cyber Defense (opens the original)
Read excerpt
BlueSTAR: Tiered Agentic Architecture for Autonomous Cyber Defense What happens when the attacker is moving faster than a human security team can respond? Modern enterprise networks are becoming increasingly difficult to defend. Security systems already generate enormous amounts of telemetry, but turning that data into the right defensive action at machine speed is still a major challenge. This video breaks down BlueSTAR, a research architecture designed to combine deterministic cybersecurity au
The Hidden Risk in LLM Supply Chains (opens the original)
Read excerpt
AI models can hide backdoors that remain completely invisible during normal use. The difficult part is not only finding these attacks, but detecting them once a poisoned model is already running in production. In this video, we break down SpecGuard, a research approach that turns a signal already produced during speculative decoding into a runtime security mechanism for LLM serving. The key idea is surprisingly simple. Modern serving systems can use a small draft model to propose tokens while a
Publishing over time
Last 90 days. Choose a month to open its work.
Recurring subjects
Named in the text we hold. One piece can cover several.
Audience
~71 subscribers
Measured Sep 19, 2026
Source's subscribers, not the number who saw an individual piece.
About this data
Counts cover the work we have indexed. Tone needs enough text and a confident classification. Excerpts and episode notes are not full articles or transcripts.
Identity or attribution wrong? Suggest a correction.