Topic · DiggingBeagle record
AI audit context completeness
An AI auditor can confuse code absent from its input with a security check absent from the system, or miss a flaw whose evidence spans components.
- Topic kind
- pattern
Definition & limits
An audit conclusion is bounded by the code and dependencies supplied to the auditor. In a system that delegates checks across components, the absence of a local check can mean either a genuine missing control or a valid check elsewhere. The distinction requires following the verification flow, including who authenticates a message, which parameters are bound, and what consumes the result.
The cross-chain bridge study reports both directions of error. Models sometimes dismissed a vulnerable flow because a required gateway component was outside the input. Elsewhere they reported an absent source-chain check in an application contract even though the upstream gateway had performed it. These examples support recording unresolved context explicitly and tracing delegated checks before treating a local observation as a whole-system verdict.
Supplying additional files alone does not guarantee correct analysis: the study separately identifies errors in protocol-specific semantics and trust boundaries. Context completeness and reasoning correctness should therefore remain separate audit questions. This is an evaluation and evidence problem, not proof that an AI auditor caused any historical bridge attack.
Examples
- A partial source-side package omits gateway logic needed to recognize an unauthorized token-sending path.
- An application delegates message authentication to its upstream gateway, so a missing local check is not automatically a vulnerability.