TLDR Sec #348: Google's PageBreak Scanner and Perplexity's Numbat Agent Security Tool
Issue #348 of TLDR Sec features Google's PageBreak scanner, an AI agent-driven web security scanner using deterministic validation, alongside Perplexity's open-source tool Numbat for monitoring agent behavior on endpoints. It also references security guidance from the UK NCSC on defending against agentic AI threats. As autonomous AI agents are increasingly deployed on endpoints and modern web applications, security teams need specialized tools to observe agent actions and filter out AI hallucinations. Combining LLM-driven vulnerability discovery with deterministic validation reduces false positives and helps organizations safely adopt autonomous workflows. Google's PageBreak project pairs LLM agents with traditional scanning tools to achieve near-zero false positives, uncovering over 500 real-world vulnerabilities through automated proof-of-exploit verification. Meanwhile, Perplexity's Go-based Numbat tool provides macOS, Linux, and Windows endpoints with local observability, pre-action blocking, and forensic reconstruction for agent activities.
## BACKGROUND
Automated security scanners powered by Large Language Models (LLMs) often struggle with high false-positive rates due to model hallucinations. Deterministic validation solves this by requiring reproducible execution proofs before confirming a vulnerability exists. Concurrently, the rise of local AI agents executing actions on behalf of users introduces endpoint security challenges that traditional antivirus tools were not built to monitor.