Eric Newcomer BrainBlog in ITOpsTimes, for Straiker
Regardless of what the calendar says, 2026 is the year of the AI agent. They are appearing all over the IT landscape, introducing new threats and vulnerabilities as they do.
Organizations rapidly build and deploy AI agents with AI driven coding tools, continuously increasing the number of deployed agents with little or no oversight and governance.
Agents automate manual tasks for operations, customer service, payment processing, analysis, research, rate setting, and their use has become a point of competition already.
It’s happening so quickly that it may seem impossible to ensure they are secure. AI agents introduce new threats and vulnerabilities wherever they are deployed and used.
The only way to discover and remediate the new threats and vulnerabilities is to fight fire with fire: use security AI agents, divided into red teams to identify the threats and vulnerabilities, and blue teams to mitigate and repair what the red teams discover, as has been traditionally done in cybersecurity.
And when the red and blue teams work so closely and rapidly together, it can seem like a purple haze descends upon that part of the landscape.
Agentic AI Threats and Vulnerabilities
AI agents raise significant new threat vectors as enterprises rapidly deploy agents throughout the enterprise, which often include third party agents in the call chain and orchestration flow. On top of that, cyber criminals also use agents to mount cyber attacks on an organization’s agents.
Agentic risks and vulnerabilities include:
- Privilege Escalation: AI agents can request performing unauthorized administrative actions and get approval.
- Indirect Prompt Injections: Attackers can embed malicious instructions into documents, websites, or emails.
- Memory Poisoning & Data Corruption: Agents form decisions based on stored memories or external tools. When adversarial agents alter stored memories or switch external tools, they compromise the agent’s decision-making logic.
- Tool Misuse: Agents are often authorized to invoke APIs and execute integrations. Insufficient controls lead to accidental data deletion, unauthorized transactions, or lateral movement across networks.
- Agentic Looping: If a task fails or is misunderstood, the agent may enter a recursive reasoning loop that exhausts compute resources, blocks operations, and racks up massive API invocation charges and token fees…
Read the article on ITOpsTimes (SDTimes sister publication) here: https://itopstimes.com/secops/secops-agents-are-red-and-blue-and-can-be-purple-too/
