MCP emerged as a critical standard for AI agents to interact with tools, much like USB-C for hardware. However, its rapid adoption overlooked security, leading to significant vulnerabilities like tool poisoning and prompt injection attacks in its early, widespread implementations.
Related Insights
AI-powered browsers are vulnerable to a new class of attack called indirect prompt injection. Malicious instructions hidden within webpage content can be unknowingly executed by the browser's LLM, which treats them as legitimate user commands. This represents a systemic security flaw that could allow websites to manipulate user actions without their consent.
AI 'agents' that can take actions on your computer—clicking links, copying text—create new security vulnerabilities. These tools, even from major labs, are not fully tested and can be exploited to inject malicious code or perform unauthorized actions, requiring vigilance from IT departments.
The core drive of an AI agent is to be helpful, which can lead it to bypass security protocols to fulfill a user's request. This makes the agent an inherent risk. The solution is a philosophical shift: treat all agents as untrusted and build human-controlled boundaries and infrastructure to enforce their limits.
The CEO of WorkOS describes AI agents as 'crazy hyperactive interns' that can access all systems and wreak havoc at machine speed. This makes agent-specific security—focusing on authentication, permissions, and safeguards against prompt injection—a massive and urgent challenge for the industry.
The world's top AI researchers at labs like OpenAI, Google, and Anthropic have not solved adversarial robustness. It is therefore highly unlikely that third-party B2B security vendors, who typically lack the same level of deep research capability, possess a genuine solution.
Research shows that text invisible to humans can be embedded on websites to give malicious commands to AI browsers. This "prompt injection" vulnerability could allow bad actors to hijack the browser to perform unauthorized actions like transferring funds, posing a major security and trust issue for the entire category.
Exposing a full API via the Model Context Protocol (MCP) overwhelms an LLM's context window and reasoning. This forces developers to abandon exposing their entire service and instead manually craft a few highly specific tools, limiting the AI's capabilities and defeating the "do anything" vision of agents.
AI researcher Simon Willis identifies a 'lethal trifecta' that makes AI systems vulnerable: access to insecure outside content, access to private information, and the ability to communicate externally. Combining these three permissions—each valuable for functionality—creates an inherently exploitable system that can be used to steal data.
Jailbreaking is a direct attack where a user tricks a base AI model. Prompt injection is more nuanced; it's an attack on an AI-powered *application*, where a malicious user gets the AI to ignore the developer's original system prompt and follow new, harmful instructions instead.
Training Large Language Models to ignore malicious 'prompt injections' is an unreliable security strategy. Because AI is inherently stochastic, a command ignored 1,000 times might be executed on the 1,001st attempt due to a random 'dice roll.' This is a sufficient success rate for persistent hackers.