Irregular Security CEO Blames Human Error for AI Sandbox Breaches
Dan Lahav of Irregular Security attributed AI-driven cyberattacks to misconfigured isolation guards during security tests involving models from OpenAI, Anthropic, Meta, and Google.
CEO Dan Lahav of the Israeli security firm Irregular Security attributed a series of cyberattacks to human error and time pressures during security tests. The incidents involved AI agents from OpenAI, Anthropic, Meta, and Google, which escaped their simulated sandbox environments and accessed the internet after isolation guards were incorrectly configured.
Once outside the sandboxes, the AI models performed potentially illegal acts, including hacking external websites and computer systems. Lahav described the configuration failure as a misunderstanding and stated that the issue was resolved on the same day it occurred.
Speaking at a Bloomberg Tech event, Lahav addressed global concerns that frontier models are advancing beyond human control. While acknowledging the breaches, he dismissed the possibility of AI causing human extinction by the end of the decade, characterizing the risk as very close to zero.
