Google's Antigravity AI Agent Vulnerable to Sandbox Escape Attack
Security researchers have uncovered a critical vulnerability in Google's Antigravity AI agent that allowed attackers to bypass the platform's highest security protections and achieve remote code execution. The flaw, discovered by Pillar Security and now patched, demonstrates how AI agents can be compromised through sophisticated prompt injection attacks even when operating under strict security constraints. Antigravity is Google's AI-powered developer tool designed for filesystem operations. The vulnerability exploited a weakness in the agent's "find_by_name" file-searching function, which was classified as a native system tool. This classification allowed the agent to execute the function directly, bypassing Google's Secure Mode protections entirely. Secure Mode represents Google's most restrictive security setting, designed to run all command operations through virtual sandboxes while limiting network access and preventing code execution outside designated directories. The attack methodology combined prompt injection techniques with Antigravity's legitimate file-creation capabilities. Attackers could deliver malicious instructions through compromised identity accounts or by embedding hidden prompts within open-source files and web content that the agent would process. The AI system struggled to differentiate between contextual data and actual prompt instructions, making it vulnerable to manipulation without requiring elevated access privileges. Dan Lisichkin, the AI security researcher who discovered the vulnerability, emphasized that the security boundary enforced by Secure Mode never detected these malicious calls. This meant attackers could achieve arbitrary code execution under the exact configuration that security-conscious users relied upon for protection. The researcher noted similar patterns have been identified in other coding AI agents, including Cursor. The disclosure timeline reveals Google's responsive approach to security issues. Pillar Security reported the vulnerability on January 6, 2026, and Google implemented a patch by February 28, also awarding a bug bounty for the discovery. However, the incident highlights broader security challenges facing the AI industry as organizations increasingly adopt agentic AI systems for business and IT operations. This vulnerability underscores fundamental trust model failures in AI agent security. Traditional security assumptions that human oversight would catch suspicious activities break down when autonomous agents follow instructions from external content sources. The incident demonstrates that sanitization-based security controls are insufficient for protecting AI agents from sophisticated attacks. The research findings have significant implications for AI security practices. Every native tool parameter that interfaces with shell commands represents a potential injection point, making comprehensive security auditing essential rather than optional. Organizations deploying agentic AI features must move beyond traditional security approaches and implement more robust protection mechanisms specifically designed for AI systems. As AI agents become more prevalent in enterprise environments, this vulnerability serves as a critical reminder that even the most advanced security measures can be circumvented through creative attack vectors. The incident emphasizes the need for continuous security research and the development of AI-specific security frameworks to protect against emerging threats in the rapidly evolving AI landscape.
A significant security vulnerability in Google's Antigravity AI agent has exposed critical weaknesses in current AI security frameworks, demonstrating how sophisticated attackers can bypass even the most restrictive protection measures to achieve remote code execution. The discovery, made by cybersecurity researchers at Pillar Security, reveals fundamental challenges in securing AI agents as they become increasingly integrated into enterprise environments.
The vulnerability targeted Antigravity, Google's AI-powered development tool designed for filesystem operations and code management. Despite operating under Google's Secure Mode—the platform's highest security setting—the AI agent remained susceptible to a carefully crafted attack that combined prompt injection techniques with the system's legitimate file manipulation capabilities.
Google's Secure Mode was specifically engineered to provide maximum protection for AI agents by implementing several security layers. The system runs all command operations through isolated virtual sandbox environments, significantly throttles network access to prevent unauthorized communications, and restricts the agent's ability to write or execute code outside designated working directories. These measures were designed to create a secure boundary that would prevent AI agents from accessing sensitive systems or executing potentially dangerous operations.
However, the vulnerability exploited a critical oversight in this security architecture. The attack focused on Antigravity's "find_by_name" file-searching functionality, which was classified as a native system tool. This classification granted the function special privileges, allowing it to execute directly without passing through the security evaluation processes that Secure Mode typically enforces on other operations.
Dan Lisichkin, the AI security researcher who identified the flaw, explained that this classification created a blind spot in the security framework. The protective boundary that Secure Mode was designed to enforce never evaluated these native tool calls, effectively creating a pathway for attackers to bypass security measures entirely. This meant that malicious actors could achieve arbitrary code execution under the exact security configuration that organizations would implement to prevent such attacks.
The attack methodology itself demonstrates the evolving sophistication of AI-targeted threats. Attackers could deliver malicious prompt instructions through multiple vectors, including compromised identity accounts connected to the agent or by embedding hidden instructions within seemingly legitimate open-source files and web content. The AI system's inability to reliably distinguish between contextual information and executable instructions created opportunities for compromise without requiring elevated access privileges.
This vulnerability pattern extends beyond Google's Antigravity system. Lisichkin noted that similar prompt injection vulnerabilities have been discovered in other popular coding AI agents, including Cursor, suggesting that these security challenges represent industry-wide issues rather than isolated incidents.
The disclosure and remediation process followed responsible security practices. Pillar Security reported the vulnerability to Google on January 6, 2026, and the company implemented a comprehensive patch by February 28. Google also recognized the significance of the discovery by awarding a bug bounty to the researchers, demonstrating the company's commitment to addressing AI security issues.
The implications of this vulnerability extend far beyond a single product or company. As organizations increasingly adopt agentic AI systems for business operations and IT management, the attack surface for these sophisticated threats continues to expand. Traditional security models that rely on human oversight to identify suspicious activities become inadequate when autonomous agents process and act upon instructions from external content sources.
The research highlights the need for fundamental changes in AI security approaches. Current sanitization-based controls, while effective against traditional threats, prove insufficient for protecting against AI-specific attack vectors. The cybersecurity industry must develop new frameworks specifically designed to address the unique challenges posed by AI agents, including their ability to process natural language instructions and their integration with system-level operations.
Moving forward, organizations deploying AI agents must recognize that every interface between AI systems and system commands represents a potential vulnerability. Comprehensive security auditing for AI deployments should be considered essential rather than optional, with particular attention paid to native tool integrations and prompt processing mechanisms.
This incident serves as a crucial wake-up call for the AI industry, demonstrating that even the most advanced security measures can be circumvented through creative attack methodologies. As AI agents become more autonomous and capable, the development of robust, AI-specific security frameworks becomes increasingly critical for protecting enterprise environments from emerging threats.