https://rexcoleman.dev/categories/research/Securing AI from the architecture up. Research, tools, and methodology for AI security. Creator of govML.https://rexcoleman.dev/images/og-default.pnghttps://rexcoleman.dev/images/og-default.pngHugoen-usTue, 31 Mar 2026 12:00:00 +0000https://rexcoleman.dev/posts/ai-security-standards-mapping/Tue, 31 Mar 2026 12:00:00 +0000https://rexcoleman.dev/posts/ai-security-standards-mapping/Cross-reference between 8 original AI security research projects and OWASP LLM Top 10, OWASP Agentic Apps, NIST AI RMF, and MITRE ATLAS. Start from any framework, find relevant research.https://rexcoleman.dev/posts/multi-agent-security/Fri, 20 Mar 2026 12:00:00 +0000https://rexcoleman.dev/posts/multi-agent-security/Simulation predicted 97% cascade poison. Real Claude agents: 60%. Topology matters (simulation said it didn't). The simulation-to-real gap changes everything.https://rexcoleman.dev/posts/llm-patch-correctness/Fri, 20 Mar 2026 12:00:00 +0000https://rexcoleman.dev/posts/llm-patch-correctness/LLM-generated patches have a 42% fix rate and 10% regression rate. SQL injection patches are net-negative — 0% fix, 50% regression.https://rexcoleman.dev/posts/llm-watermark-robustness/Fri, 20 Mar 2026 11:00:00 +0000https://rexcoleman.dev/posts/llm-watermark-robustness/Cross-model paraphrasing drops watermark detection from 100% to 60% in one pass, then plateaus at 40% after 10 passes. Kirchenbauer green-list watermarks are partially robust — but not enough for adversarial settings.https://rexcoleman.dev/posts/agent-semantic-resistance/Fri, 20 Mar 2026 10:00:00 +0000https://rexcoleman.dev/posts/agent-semantic-resistance/First taxonomy of why real LLM agents resist cascade poisoning — and which attacks bypass each resistance pattern.https://rexcoleman.dev/posts/agent-vuln-triage/Fri, 20 Mar 2026 10:00:00 +0000https://rexcoleman.dev/posts/agent-vuln-triage/An LLM agent achieves 92% precision@10 on vulnerability triage but underperforms EPSS (100%). The ensemble reaches 98% with lower variance.https://rexcoleman.dev/posts/verified-delegation-protocol/Thu, 19 Mar 2026 22:00:00 +0000https://rexcoleman.dev/posts/verified-delegation-protocol/We proposed a 3-layer defense for multi-agent cascade. Real agent experiments refuted 5 of 7 hypotheses. The simulation was wrong by 48 percentage points.https://rexcoleman.dev/posts/financial-anomaly-detection/Thu, 19 Mar 2026 00:00:00 +0000https://rexcoleman.dev/posts/financial-anomaly-detection/CFA-informed rule-based scoring achieves 0.898 AUC on its own, and 8 of the top 20 predictive features come from domain expertise, not raw data.https://rexcoleman.dev/posts/pqc-migration-analyzer/Thu, 19 Mar 2026 00:00:00 +0000https://rexcoleman.dev/posts/pqc-migration-analyzer/70% of the crypto in your codebase isn't yours to change — and classical exploit risk matters more than quantum vulnerability for deciding what to fix first.https://rexcoleman.dev/posts/rl-agent-vulnerability/Mon, 16 Mar 2026 22:00:00 +0000https://rexcoleman.dev/posts/rl-agent-vulnerability/I trained RL agents on security tasks, then attacked their reward functions, observations, and policies. Observation perturbation is 20-50x more effective than reward poisoning. Policy extraction achieves 72% agreement with just 500 queries.https://rexcoleman.dev/posts/model-fingerprinting/Mon, 16 Mar 2026 00:00:00 +0000https://rexcoleman.dev/posts/model-fingerprinting/How do you know a model downloaded from Hugging Face hasn't been backdoored? I built a behavioral fingerprinting system that uses unsupervised anomaly detection to answer that question.https://rexcoleman.dev/posts/agent-redteam/Mon, 16 Mar 2026 00:00:00 +0000https://rexcoleman.dev/posts/agent-redteam/19 attack scenarios against LangChain and CrewAI agents. 100% success with reasoning chain hijacking. 7 attack classes systematized — 5 not in OWASP or MITRE ATLAS.https://rexcoleman.dev/posts/adversarial-control-analysis/Mon, 16 Mar 2026 00:00:00 +0000https://rexcoleman.dev/posts/adversarial-control-analysis/The same security principle — classify inputs by who controls them — works across network IDS, vulnerability management, AI agents, post-quantum crypto, fraud detection, and AI supply chains.https://rexcoleman.dev/posts/adversarial-ids/Sat, 14 Mar 2026 00:00:00 +0000https://rexcoleman.dev/posts/adversarial-ids/I built and red-teamed an ML-based intrusion detection system. The key finding: which features an attacker controls matters more than which model you choose.https://rexcoleman.dev/posts/cvss-gets-it-wrong/Sat, 14 Mar 2026 00:00:00 +0000https://rexcoleman.dev/posts/cvss-gets-it-wrong/I trained an ML model on 338,000 CVEs to find out what actually predicts exploitation. CVSS scores severity. Attackers measure opportunity. The model reveals what they look for.