Untrusted input reaches an LLM call
What a llm-call-injection finding of fallow security means, why it matters and how to fix it.
What it reports
Catalogue security candidate for CWE-1427.
Security explains how fallow finds it.
Why it matters
Input from the request goes into the prompt. An attacker can write instructions that change what the model does, for example to leak the system prompt or call tools.
Example
openai.chat.completions.create({ messages: [{ role: 'user', content: req.body.q }] });
Fix
Review the candidate trace before acting. Confirm attacker control, missing sanitization, and reachable runtime context, then fix with the category-appropriate safer API or add a reviewed suppression.
Facts
| Fact | Value |
|---|---|
| Issue type | llm-call-injection |
| Rule id in SARIF output | security/llm-call-injection |
| Command | fallow security |
| Default severity | off |
| Suppression comment | // fallow-ignore-next-line security-sink |
Fixable with fallow fix | No |
| Opt-in | Yes |
Note. Tainted-sink catalogue category; the security-sink suppression token covers every category.