Brief

• Akamai launched the beta versions of Agent Sandbox and Akamai Inference on October 27th, adding an isolation layer based on WebAssembly to the code written and run for the AI proxy.
• A survey conducted in 2026 among 2,930 security executives by Palo Alto Networks found that on average, each individual corresponded to 109 machine identities, of which 79 were AI proxies.
• Palo Alto, Cisco, CrowdStrike, Okta, and Zscaler have all acquired or launched AI proxy security products this year.

Cloud service provider BJW_KEEP_00003__ launched a sandbox for BJW_KEEP_00004__ agents on Tuesday. This is a closed environment where agent-written code can run without accessing other parts of the company's systems. BJW_KEEP_00001__ is currently in beta, and BJW_KEEP_00002__, a service for running the open-source BJW_KEEP_00007__ model on the BJW_KEEP_00006__ cloud, was also launched simultaneously.

This release is not on a large scale, but the issues behind it are not minor. Software is no longer just about answering questions; it can log into accounts, write code, transfer funds, and act autonomously, and each of these capabilities can be misused by hackers or by the software itself.

This is what's known as proxy-based AI ( agentic AI ), which is taking over the internet. A survey conducted by Palo Alto Networks in 2026 among 2,930 security responsibles found that on average, each person corresponds to 109 machine identities, which are the login information and keys used by software to access other software. Among these, 79 are AI proxies.

Proxy-based AI brings many benefits, but it also comes with serious risks, especially when these proxies have too much freedom or are intelligent enough to bypass their own limitations.

For example, on July 16th, an important platform for open-source AI models, Hugging Face, disclosed that it had been hacked. Five days later, OpenAI confirmed that its own models were also involved in this incident. The research institution METR subsequently investigated the incident and found that approximately 1,200 OpenAI proxies were acting together on a hidden message board, with about 700 of them participating in the attack.

The results show that these proxies escaped the sandbox and exploited an previously unknown software vulnerability to gain access to the real system, combining it with stolen credentials. OpenAI referred to this as a "warning shot." The company stated that during the testing period, a monitoring tool used to read the written reasoning process of the model was not in operation.

In simple terms, AI security often ultimately evolves into a tug-of-war between “alignment” and “development” – the former is responsible for ensuring AI security, while the latter is responsible for ensuring AI strength. Both parties hope that the other will play a secondary role in order to achieve their own goals.

Akamai hopes to address precisely this issue. Akamai, the Senior Vice President of Cloud Technology, stated in a press release: "When companies deploy AI into production, they should not have to choose between speed and control."

Some teams hope to launch a AI feature on the hosting platform this quarter. Other teams, on the other hand, need dedicated GPU to run cutting-edge open-source models in their own way. Every team that runs proxies needs to establish boundaries around the code they write for these proxies, he said.

Hijacking is another concern. In July, Zscaler discovered that web pages contained instructions specifically written for AI's proxy, a technique known as indirect hint injection. A forged document page required the proxy to pay a "license fee" of $3, which is approximately 0.0012 Ethereum, to be transferred to a hardcoded wallet; among the 26 models tested by Zscaler, 4 followed this instruction.

Going back to May, cybersecurity company Bugcrowd stated that the number of submissions in March more than quadrupled within three weeks, with the majority of them being fake. HackerOne and Nextcloud suspended their paid projects in April.

Machines can also discover real vulnerabilities. Mozilla indicates that earlier versions of Claude Mythos helped to identify 271 vulnerabilities in Firefox, all of which were fixed within a week. The massive influx of both real and fake discoveries is precisely what security companies are now competing to manage.

Palo Alto completed the acquisition of CyberArk on February 11th in order to extend privileged access control to AI proxies. In April, the company acquired Koi, which is responsible for protecting AI encoding proxies running on employees' machines. Cisco announced plans to acquire Astrix Security in May; this company is in charge of managing the keys and tokens used by the proxies, and the transaction was completed on June 29th.

The approach of Akamai is to set up barriers around the code itself. The company's sandbox uses WebAssembly – a format that runs programs in strictly isolated environments – to limit the computing power, memory, and external connections of the proxies. According to Alexander of Akamai, every team that operates proxies “needs to establish a boundary around the code they write for those proxies.”

According to the survey by Palo Alto, institutions expect the growth of AI agents to be 85% in the next 12 months, while the growth rate for human identities is 56%.