Advanced Security Topics · Fall 2026

Project Ideas

Research themes for LLM-based security projects

Projects should investigate a focused research question involving LLMs and security. The themes below are intentionally broad: students may propose ambitious systems, careful empirical studies, new evaluation methods, or smaller but technically sharp projects within any of these directions.

1. Verifying LLM-Generated Security Fixes

Core question: How can we determine whether an LLM-generated security patch is actually correct and secure?

Check whether a generated patch truly eliminates the vulnerability rather than merely passing existing tests.
Automatically generate adversarial regression tests for LLM-produced patches.
Use execution, static analysis, symbolic reasoning, or formal verification to validate patches.
Detect whether a patch introduces new vulnerabilities or breaks functionality elsewhere.
Compare generate-then-check pipelines with verifier-guided patch generation.
Characterize common failure modes of LLM-generated security fixes.

2. Optimizing LLM Vulnerability-Discovery Trajectories

Core question: What is the right search strategy for an LLM acting as an autonomous vulnerability researcher?

Learn search policies from successful and failed vulnerability-discovery traces.
Decide which file, function, call path, or hypothesis the model should inspect next.
Study when an agent should deepen one hypothesis, branch to alternatives, or backtrack.
Optimize search under fixed token, time, or tool-call budgets.
Define useful intermediate notions of progress beyond a binary success/failure outcome.
Compare breadth-first, depth-first, uncertainty-guided, and learned search strategies.

3. LLM Security Reasoning and Tool Use

Core question: When should an LLM rely on latent security reasoning, and when should it query an external tool or environment?

Security tasks often require a mixture of internal reasoning and external semantic feedback. A security agent may choose among actions such as running code, debugging, fuzzing, invoking a static analyzer, querying an SMT solver, running tests, inspecting dependencies, or searching documentation.

Learn policies that decide when a tool call is worth its cost.
Measure when latent reasoning is sufficient and when external verification is essential.
Study whether uncertainty or confidence can predict useful tool calls.
Compare tool-use policies across vulnerability discovery, patching, exploit development, malware analysis, reverse engineering, incident response, and secure code review.
Investigate whether different security tasks require different sequences of tools.
Study cases where tool feedback hurts by distracting or misleading the model.
Frame tool selection as value of information, adaptive computation, or experimental design.

4. Measuring Genuine LLM Security Reasoning

Core question: How can we distinguish genuine security reasoning from memorization, benchmark contamination, pattern matching, or lucky search?

Build temporal holdout benchmarks using vulnerabilities disclosed after likely model training cutoffs.
Apply semantics-preserving code transformations and test whether model conclusions remain stable.
Separate vulnerability localization, root-cause reasoning, trigger construction, exploit generation, and repair.
Test how performance changes as progressively more hints, patches, crashes, or CVE information are revealed.
Develop evaluation methods for previously unseen vulnerabilities rather than recycled benchmark tasks.
Measure robustness of security reasoning across codebases, languages, and repository structure.

These are starting points, not fixed assignments. Students are encouraged to propose their own question within the broader goal of understanding and improving LLM-based security reasoning.