Full copies are stored in the repository and served from this site — 4 are marked as anchor reading by the research.
Ohio State/NYU. CodeLlama-13B synthesizes Python-based generation fuzzers. Up to 434.8% more code coverage than second-best SOTA on cvc5 within 10 min. Fuzzer space model (lattice structure) contributes up to 62.5% to effectiveness.
Four-question framework (Adam Shostack): What are we working on? / What can go wrong? / What are we going to do about it? / Did we do a good enough job? STRIDE categorization. anchor for agent workflow design.
A03:2025 Software Supply Chain Failures newly elevated to top-10. A01:2025 Broken Access Control remains #1. Based on NVD data (~175,000 CVEs, 643 CWEs), CVSSv2/v3, incidence rates. Data covers 2021-2024.
UCSB paper. Three-phase pipeline: static analysis→LLM harness synthesis→symbolic execution→concrete validation. 379 vulns vs 12 by Claude Code alone across 10 C/C++ projects (6.8M LoC). Ablation: removing any component collapses results. anchor for agent architecture.
Five-stage lifecycle defense-in-depth: initialization → input processing → memory → decision-making → execution. Plugin-native prototype on OpenClaw. Cross-layer threat propagation interception.
Berkeley team (Dawn Song). Black-box fuzzing framework for indirect prompt injection. 71% success vs o3-mini agents, 70% vs GPT-4o on AgentDojo/VWA-adv. MCTS-based seed selection.
Empirical study: 7 LLMs vs SAST tools for security code review. LLMs significantly outperform SAST. DeepSeek-R1 best, then GPT-4. CoT + commit message prompt optimal.
Fully autonomous LLM-powered SSH pentesting agent. Multi-LLM (GPT-o3-mini planning, LLaMA 3.1 command formulation, GPT-4o summarization). 60% success vs ShelLM defender, 57.58% vs OverTheWire Bandit, avg <5 actions when winning.
16-author paper (Invariant Labs, IBM, EPFL, ETH Zurich, Google, Microsoft, Swisscom, Kyutai). Systematizes prompt injection defense patterns for LLM agents. Foundational reading for agent security architecture.
Springer 2026. GPT-4o correctness on sorting benchmarks: 35% → 67.50% with fuzzing feedback loop. Demonstrates LLM+fuzzing feedback improves program synthesis correctness. (Note: 'dramatically improves' framing overgeneralized — verified as scoped result only.)
NUS paper. Achieves concolic execution effects without symbolic constraint solving. Program slicing + LLM input modification suggestions. 31-59% more branches covered. (Note: speed claims refuted in verification.)
Google DeepMind's adversarial evaluation framework for Gemini. Adaptive attack suite run continuously against past/current/future model versions. Practical lessons on adversarial robustness testing.
Standardized benchmark for prompt injection attack/defense evaluation. Adaptive testing with white-box and black-box optimization. Shows current defenses remain susceptible.
Layered security framework: W3C DIDs for identity, NIST PQC for communication, Halo2 ZKP for policy compliance. Formal adversary model extending Dolev-Yao. 0% success across 20,000 attack trials in simulation.
NOVA system: multi-agent, multi-model, 3,915 OSS projects → 14,090 confirmed vulns in 2 months, 99.4% unreported, 40% high/critical. Pipeline: review history → read code → identify candidates → create PoC → validate in clean env → patch candidate → disclosure report.