Simulate adversarial user queries to identify vulnerabilities, policy breaches, or toxic outputs.
Deliberately attempts to make the system misbehave: bypass safety constraints, disclose system instructions, invoke tools it should not, or produce prohibited content. Adversarial testing before an adversary does it.
Any system accepting untrusted input, and every agentic system with write-capable tools regardless of who the users are.
Closed internal pipelines with no free-form input path, where the adversarial surface is genuinely absent.
Findings must be tracked to remediation like security findings, not filed as observations. Red-team results with no closure process are theatre.
Testing the prompt surface and ignoring the tool surface. In agentic systems the serious vulnerability is what the agent can do, not what it can say.
An agentic exception investigator tested with instructions embedded inside an uploaded evidence document, checking whether it treats document content as data rather than as direction.
The source material did not expand this solution into components, and the row says so in its own notes. Expanding it here would be authorship, not research.
Two different absences share this shape. Foundational solutions get built whatever the domain, so no domain links them; the rest are solutions this domain genuinely does not reach for. v_ai_solutions_unlinked separates the two.