Four Recent Studies Reveal Critical Security Gaps in AI Agents

Key Points

• Manifold Security found a Chrome extension can fake user clicks, letting Anthropic's Claude read Gmail and other personal data silently.
• Researchers showed a single malicious email can plant a persistent false memory in an AI assistant, influencing future sessions.
• Semgrep team demonstrated cheap model poisoning that embeds hidden code vulnerabilities into open‑weight AI models.
• PromptArmor tracked rapid changes in AI connectors, with 931 of 2,517 altered in six weeks, expanding attack surface dramatically.
• All four studies highlight the same core weakness: surrounding infrastructure that lets agents act on user data without proper checks.