The original headline is "Code vs Judgment: What Adversarial AI Testing Revealed About Verification Failures"

Short summary

An engineer describes using Grok as an adversarial critic against their own AI agent work, discovering that the real failure line isn't human vs machine but code vs judgment. Code-based hooks enforce rules without bargaining, while judgment-based checks are soft on both sides of the loop — the agent rationalizes just like the human does. The key insight: decorrelating your critic from your work surface catches blind spots that shared assumptions miss.

• •Used Grok adversarially to attack own AI agent work — surfaced human-layer failures
• •Code-based guardrails remove the right to bargain; judgment-based checks are soft on both sides
• •Verification needs different resistance types for different failure modes

Generated with AI, which can make mistakes.