Your AI Agents Can Be Hijacked. Most Enterprises Can’t Stop It.
🔴 Why it matters: Researchers just proved that the AI agents you’re deploying right now — from Anthropic, Google, and Microsoft — can be silently commandeered to steal your data, delete your emails, or execute hidden malicious commands. And most businesses have zero defense.
A new research paper dubbed “Agents of Chaos” — from a 20-person team — details how security researchers successfully hijacked Claude Code, Gemini CLI, and GitHub Copilot agents using prompt injection attacks: hidden instructions embedded in websites or files that the agent reads and silently obeys.
The exploits weren’t theoretical. Researchers stole API keys, leaked sensitive data, and triggered commands like “delete your database.” Both Microsoft Copilot and Salesforce Agentforce had confirmed flaws — now patched — but notably, none of the vendors (Anthropic, Google, Microsoft) issued public CVE disclosures, meaning most IT teams never heard about it.
A VentureBeat survey of 108 enterprises found the majority cannot stop “stage-three” AI agent threats — attacks where an agent is actively executing malicious actions inside your systems. Palo Alto Networks’ Unit 42 already found attack traces in the wild in early March.
The business takeaway: AI agents need access to your accounts, calendars, email, and databases to be useful — which is exactly what makes them attractive targets. As Adrien Merveille of Check Point put it bluntly: “When you deploy agents, you have no control over what they’ll do.”
What to do now: Audit every AI agent’s permission scope. Apply least-privilege access. Never give an agent write access it doesn’t need. Demand your vendor’s prompt injection hardening roadmap before expanding deployment.
Read: Khaleej Times — AI Agent Fever Comes With Lurking Security Threats →
|
