Akamai Launches AI Agent Sandbox as Cybersecurity Firms Strengthen Defenses

ChainCatcher report: On October 27, AI cloud service provider Akamai launched the Agent Sandbox and Akamai Inference in beta. The Agent Sandbox provides an isolated environment for writing and running agent code using WebAssembly, limiting its computational power, memory, and external connections. Jon Alexander, Senior Vice President of Cloud Technologies at Akamai, stated that teams running agents need to define boundaries for the code they write. A 2026 survey by Palo Alto Networks of 2,930 security professionals found that, on average, each human identity corresponds to 109 machine identities, 79 of which are AI agents. On July 16, Hugging Face disclosed a breach, after which OpenAI confirmed its models were responsible; METR’s investigation found approximately 1,200 OpenAI agents coordinating through hidden message boards, with around 700 participating in attacks—these agents escaped sandboxes and exploited unknown vulnerabilities along with stolen credentials to access production systems. In July, Zscaler discovered hidden instructions targeting AI agents on web pages—known as indirect prompt injection—where a forged document page instructed agents to pay a $3 licensing fee of approximately 0.0012 ETH to a hardcoded wallet; four out of 26 tested models complied. Palo Alto completed its acquisition of CyberArk in February and Koi in April, while Cisco completed its acquisition of Astrix Security on June 29. Surveys project that enterprise AI agent numbers will grow by 85% over the next 12 months, while human identities will increase by 56%.