TrendAI™ presented two sessions at the [un]prompted 2026 conference in San Francisco. The first, by Sean Park, demonstrated how AI-driven KYC pipelines can be exploited by embedding malicious instructions inside identity documents like passports. Using a real-world stack with FastAPI, Claude Code, and a SQLite MCP backend, the team showed how hidden injects can cause AI agents to leak customer records. They ran 2,600 automated tests across 13 models to identify high-success-rate attack vectors. The second session introduced FENRIR, a multi-stage automated vulnerability discovery system combining CodeQL, Semgrep, YARA-X, SpotBugs, and LLM reasoning to find zero-days in AI and MCP ecosystems. FENRIR eliminates over 90% of false positives before human review and has already produced 60+ published CVEs, 100+ vulnerabilities in pre-disclosure, and 3,000+ findings queued for review.
Sort: