Silent LLM Compliance Poses Greater Security Risk Than Loud Failures, Audit Finds

A week-long security review of LLM-powered applications revealed that the most dangerous failures are not blocked prompts or thrown exceptions, but quiet, undetected compliance with requests the model should have refused. Dubbed 'soft failures,' these incidents return valid, well-formed responses that logs and monitors treat as successful, making them nearly invisible to standard testing. During one audit of a banking-style assistant, a politely worded, cross-language probe bypassed restrictions that an aggressive jailbreak attempt could not. The reviewer recommends detecting soft failures by comparing a model's capability before and after a suspected injection, rather than scanning response text for suspicious content. Running a focused battery of 8–15 adversarial probes against a live system prompt is suggested as a practical first step to identify whether such vulnerabilities exist.