Bypasses / Prompt guardrail or safety filter
Prompt guardrail or safety filter
8 cases that got past this protection.
- PraisonAI: Prompt-injection defense blocks only when 3+ detector families fire simultaneously; realistic single-vector injections pass through unblocked The prompt-injection scanner blocks only when several detector families fire together, so a single-pattern override instruction passes through unblocked.
- PraisonAI: Prompt Injection Defense Bypassed for HIGH-Level Threats The injection scanner only blocks when several checks fire, so single-pattern override attempts are logged but still forwarded to the model.
- vLLM: Processing differential in multi-channel audio downmixing enables hidden-input/moderation bypass for audio models Content hidden in unused audio channels is heard by the model after mono downmixing but not by human listeners, evading moderation.
- BerriAI litellm has Security Feature Bypass in BannedKeywords and AzureContentSafety Guardrails via call_type Mismatch on Async Endpoints BannedKeywords and Azure content safety guardrails are skipped on async endpoints because the hook's call type does not match, letting prompts pass unfiltered.
- FastChat has a Content Moderation Bypass via Arena Side-by-Side Views Requests to the arena side-by-side handler take a control flow path that skips content moderation, because the earlier fix missed three related files.
- Flowise: Remote code execution vulnerability in AirtableAgent.ts caused by lack of input verification when using `Pandas`. User questions are inserted unsanitised into the prompt, so a crafted prompt makes the model generate Python that the Pyodide runtime executes.
- n8n has a Guardrail Node Bypass Crafted user input in an n8n workflow gets past the default instructions of the Guardrail node, defeating its safety checks.
- PandasAI interactive prompt function Remote Code Execution (RCE) Crafted natural-language prompts can make the LLM emit Python that PandasAI runs without adequate checks, giving the attacker code execution on the host.