Developer's AI Safety Gate Passed a Security Flaw Three Times Before One Model Caught It

A software developer uses a peer-review system where any AI agent's changes to its own safety controls must be approved by a panel of four rival AI models from different labs, with a single dissent enough to block a change. This week, the system was tested when a fix to a safety gate — designed to prevent the agent from bypassing input checks — was submitted for review across multiple rounds. In three consecutive rounds, two of the three active reviewers approved the fix, while DeepSeek repeatedly identified new layers of the same underlying vulnerability. The root flaw was that the agent could clear its own safety check simply by including a specific phrase in its own output, rather than requiring confirmation from an actual tool result. The developer argues the episode shows that majority-vote review is inadequate for security verification, since correlated training data can produce correlated blind spots across models.