SNOWFLAKE
DATA BREACHES
SNOWFLAKE
HEADQUARTERS: MENLO PARK, CALIFORNIA, US
NUMBER OF EMPLOYEES: 9,000 + MARKET CAP: US $ 89BN NUMBER OF COUNTRIES: 24 XXX
Constraining the autonomous agent To combat these evolving threats, Mayank’ s team has rolled out targeted defence mechanisms. The initial battleground is the data layer, where malicious prompts can be neutralised before the AI even reads them.
“ One of the things we did is we announced a suite around indirect prompt injection protection, which works without adding any latency,” Mayank explains.“ It’ s uses an LLM behind the scenes to figure out what may be going on and understand if someone’ s trying to trick you.”
But protecting the AI from deceptive data is only the first step. The second, more complex challenge involves securing the harness – the digital framework that controls what an agent is actually authorised to do when acting on a user’ s behalf.
To prevent autonomous agents from running amok, Mayank emphasises the need for strict guardrails rather than polite requests.“ We released something called Agent Identity,” he says.“ Agents are different to the people that own and control them so they should be constrained in what they can do.
cybermagazine. com 99