AI/ML Internship: Self-Improving AI Agents for Cybersecurity

Summary

AI agents have reached the point where they find real vulnerabilities in software. How well they do depends heavily on the harness around the model: its tools, scaffolding, model choice, memory, and orchestration. Hand-tuning these harnesses does not scale, and agents that improve their own harnesses are still largely unexplored especially for security-related tasks. In this role, you will contribute to this effort: building on existing work towards creating agents and methods that design, evaluate, and improve their own harness. Apple's Security Engineering & Architecture organization is responsible for the security of every Apple product. AI is rapidly changing how cybersecurity work gets done. When it comes to securing more than a billion devices running the world's most sophisticated operating systems, that means ensuring that we make full use of AI's capabilities while maintaining high-quality standards. This internship focuses on developing methods for automatically improving AI agents for security tasks. During this internship, you will work with some of the best security researchers in the world to apply AI to real-world security problems and to develop new AI methods. We are looking for students in AI/ML with an interest in agentic AI who want to take on the hard, open problems that cybersecurity offers. No prior experience in cybersecurity is required.

Minimum Qualifications

Currently doing a degree in one of the following fields: machine learning, AI, software engineering, computer science, security, or a related field. Experience building, experimenting with, and evaluating agentic AI systems (e.g., with LangChain, LangGraph, OpenHands SDK, Pi). Strong software engineering skills in Python, and experience with agentic coding. Autonomous drive, a collaborative mindset, and a proactive approach to problem-solving.

Preferred Qualifications

Experience with building and evaluating agentic software-related tasks (e.g., vulnerability research), agentic self-improvement, or memory. Knowledge of statistics and black-box optimization methods.

Description

Security tasks are a distinctive setting for agentic self-improvement. (a) Success is highly verifiable but rare. (b) The search space is vast and the horizons are extremely long. (c) They are grounded in real-life codebases that constantly evolve. These properties make security both a demanding test and a rich source of signal for continuously improving agents automatically. While base models provide the basis of an agent's capability, the harness around them has a large effect on many of the required attributes. The goal of this internship is to experiment with new self-improvement methods on real-world security problems. You will study how to automatically improve harnesses on cost, performance, robustness, and how to feed-in signals from the software development lifecycle into that process. Novel ideas will be required to tackle these challenges. Code and the agent's environment constantly evolve: fixes for past issues may not lead to future ones, and the harness must keep evolving. Because agents are non-deterministic and successes are sparse, telling signal from noise is both critical and hard. Solving these challenges calls for novel techniques that draw on state-of-the-art LLM research, active learning, black-box optimization, software engineering practice, and program analysis.