What it asks
Limit what an agent can do on its own: minimal tools and permissions, human approval for high-impact actions, authorization enforced outside the model, and logging.
Compliance mapping
OWASP LLM06, EU AI Act art. 14 and LGPD art. 46.
Where a runtime decision on each agent action helps with three common requirements, and what stays with you. Vulnify has no certification and does not make you compliant. This page is not legal advice.
OWASP Top 10 for LLM Applications: LLM06 Excessive Agency
Where Vulnify helps
Every action is decided outside the model before it runs: ALLOW, REVIEW or BLOCK with a 0–100 score and reasons. REVIEW waits for a different person. An agent changing its own permissions is flagged. Every decision lands in a hash-chained audit log.
What stays with you
Scoping the agent’s tools and credentials in the first place, and choosing which actions need a human.
EU AI Act, Article 14: human oversight
What it asks
High-risk AI systems must be designed so people can oversee them: understand the output, decide not to use it, override it, and stop the system.
Where Vulnify helps
REVIEW puts a named person between the agent and the action, in the app or in Slack. Reasons explain each decision. BLOCK and pausing or disabling an agent stop it. Monitor mode lets you watch before you enforce.
What stays with you
Deciding whether your system is high-risk under the Act, assigning competent overseers, and the rest of the obligations (risk management, documentation, conformity).
LGPD, Art. 46: security measures
What it asks
Controllers and processors must adopt technical and administrative measures to protect personal data from unauthorized access and from accidental or unlawful destruction, loss, alteration, communication or disclosure.
Where Vulnify helps
Personal data going out, bulk exports and deletes score higher and can be blocked or sent to review. Content is scanned for personal data, with extra weight for sensitive data under art. 11 (the content itself is not kept). Free LGPD policy templates. Audit export in JSON or CSV with a SHA-256 checksum.
What stays with you
Your legal basis, records of processing, data-subject requests, the encarregado, and every control that is not about agent actions.
The scoring behind these decisions is public.