Hallucinated Alert
In scenarios such as security auditing, the phenomenon in which an AI model fabricates non-existent issues or call chains, presenting seemingly plausible false alerts in the form of technical terminology.
Timeline (last 90 days)
The model may give different conclusions when asked twice about the same piece of code, and may also give a plausible-sounding description of a problem that does not exist (hallucinated alerts).
Means of suppressing hallucinated alerts include: having the model cite specific line numbers and code snippets, cross-validating by asking a second question about the same code from a different angle, and allowing "cannot confirm" as a permitted output.
All Facts (2)
The model may give different conclusions when asked twice about the same piece of code, and may also give a plausible-sounding description of a problem that does not exist (hallucinated alerts).
50%UnverifiedMeans of suppressing hallucinated alerts include: having the model cite specific line numbers and code snippets, cross-validating by asking a second question about the same code from a different angle, and allowing "cannot confirm" as a permitted output.
50%