'The agent itself has become its own entity to secure': Bitdefender's new free Mac tool goes after flaws that let attackers fool AI models

Bitdefender has introduced a new free security tool called AI Guardian for Mac users, designed to safeguard autonomous software agents from potential threats. This innovative application operates in the background, meticulously scrutinizing the actions requested by agents before they are executed. AI Guardian meticulously checks these operations against user-defined rules, providing a green light, a warning, or a complete block as necessary.

It records each decision for future review, enabling users to examine the details of each operation and ensuring transparency.

The tool is currently compatible with Claude Code 2.1.121 and OpenClaw 2026.6.6, and it operates locally, although it can utilize Bitdefender’s cloud infrastructure for certain services like website reputation checks. AI Guardian is equipped to identify attempts to manipulate agents through hidden instructions and prevent these operations from proceeding without authorization. It also examines Model Context Protocol tools, blocking any suspicious or altered components that could compromise the agent.

Security checks are performed on the supported agent skills before execution, and exposed API credentials can trigger restrictions around protected resources. The software can block agents from accessing sensitive resources such as SSH keys and system credentials without proper authorization. These measures are primarily intended for developers, technical practitioners, and those who rely on agents for coding or operational tasks.

The company emphasizes that every decision made by AI Guardian can be reviewed through a detailed auditable record, providing full transparency into how individual requests are handled.

Research conducted by Bitdefender reveals the urgent need for such a tool. In testing 20 AI agents against over 1,300 tool poisoning attempts, malicious attacks were successful in 36.5% of cases. In one instance, a model was manipulated in 72.8% of attempts, highlighting the significant risks associated with agent capabilities beyond simple chatbot interactions.

A separate study reported more than 1.2 million exposed AI service secrets during 2025, marking an 81% increase from the previous year. Additionally, over 24,000 credentials were exposed through publicly available Model Configuration Protocol (MCP) configurations during the same period.

Ciprian Istrate, senior vice president of operations at Bitdefender, emphasized the importance of securing AI agents. "AI agents are becoming a direct extension of the users who rely on them, inheriting the same security risks that come with that role," he stated. "This rapid shift means security can no longer stop at protecting the human alone; the agent itself has become its own entity to secure."

AI Guardian is currently available as a public beta, with plans to expand its support to other operating systems in the future. The software is currently available in English and for macOS users.