Rahul Suresh Babu

Despite perfect tool selection, tool-augmented agents still make critical mistakes by binding to the wrong entities in 25% of cases, revealing a hidden layer of reliability issues.

Effect forgery poses a greater threat to LLM safety than risk label tampering, revealing a critical vulnerability in tool contract integrity.

GIST-CMTF slashes wrong-goal execution rates from 19.4% to just 2.5%, proving that validating user goals is crucial for effective tool-augmented agents.

CMTF slashes tool exposure from 100 to just one per step, cutting token costs by 90% while maintaining task success rates.