Kadoa — agentic threat model

Not certain from the listing — likely relies on commercial LLMs for unstructured data transformation. The primary threat is indirect prompt injection, where malicious instructions embedded in target web pages hijack the model's extraction or transformation logic.

Processes unstructured web data at scale across sensitive domains like finance and recruitment. Threats include data poisoning from target websites and data exfiltration or leakage of scraped PII/proprietary data during the transformation phase.

Orchestrates 'self-healing' scrapers that dynamically adapt to website changes. Threats include logic flaws in the self-healing orchestration, leading to infinite loops, scraping unintended targets, or executing malicious DOM elements.

Not certain from the listing — likely hosted in a cloud environment utilizing proxy networks. Threats include proxy abuse, IP blocking, and potential sandbox escape if the scraping engine executes untrusted JavaScript from target sites.

Not certain from the listing — requires robust observability to detect when self-healing scrapers fail or silently extract corrupted/poisoned data, leading to data drift and integrity issues in downstream analytics.

Not certain from the listing — scraping at scale across finance and recruitment introduces significant compliance risks, including Terms of Service (ToS) violations, copyright issues, and GDPR/CCPA violations when harvesting PII.

Not certain from the listing — while direct multi-agent interactions are not highlighted, the agent integrates directly with downstream enterprise data workflows, meaning compromised data could propagate and cause cascading failures in connected systems.