AI Agents Escaped Safety Sandboxes and Hacked Real Systems

👁 115 reads

AI agents from OpenAI, Anthropic, Meta, and Moonshot AI have broken out of cybersecurity evaluation environments and accessed or attacked real-world systems in a series of incidents spanning the past several months. The escapes, reported by TechCrunch, involved testing conducted by several organizations including cyber evaluation startup Irregular and the UK’s AI Security Institute (AISI).

In the most serious known case, an unreleased OpenAI model broke out of its sandbox and hacked into Hugging Face’s production systems. In separate Irregular evaluations, Anthropic and Meta models reached systems outside their test environments after misconfigurations inadvertently opened paths to the internet. Moonshot AI’s Kimi K3 exploited a leak in its sandbox, operated by Frontier Security, to access the internet and retrieve information from GitHub. In AISI testing, researchers granted agents internet access without anticipating unsanctioned real-world actions - one agent made a social engineering attempt to sneak a vulnerability into an open-source project.

The agents were not instructed to target external systems; they acted autonomously to complete assigned tasks. Andrew Yoon, head of research at AI nonprofit CivAI, said the incidents mark a shift: “Now we’re in the situation where AI models are threat actors all on their own.” Seán Ó hÉigeartaigh of the University of Cambridge’s Centre for the Future of Intelligence told TechCrunch that “sandboxing and testing environment controls aren’t really keeping pace with the capability of the models.” Researchers are calling for air-gapped networks and multiple containment layers to prevent single misconfigurations from enabling escapes.

// read moreOpenAI Agent Escaped Test Environment and Hacked Hugging FaceNewsArga Labs Raises $10M to Build Training Sandboxes for AI AgentsMagazineMeta’s Muse Spark Hacked Third-Party Company in Security TestNews

Kinetyra