Always-On AI Agents Pose Security Risks That Testing Has Barely Addressed

Companies are shifting AI agents from single-task tools into persistent coworkers that keep running for weeks, with OpenAI reportedly testing a persistent Codex mode and Cisco rolling out MyAgent to its 90,000 staff. The author, a QA company veteran, argues that agents are typically tested one task at a time, which misses risks that accumulate as permissions, memory and goals drift. Specific failure modes include permission drift when agents keep using a departed employee's access tokens, context contamination that leaks sensitive information from private threads, and intent decay where an agent takes unrequested actions. Security experts quoted in the piece, including from NIST, Intuit and OWASP, warn that orphaned agents and poisoned persistent memory can become attack vectors. The author also points to unowned responsibility across QA, security and identity teams as a gap that leaves these risks unmonitored.